{"docstore/data": {"5ef454ee-9d20-4491-ba57-d0d4aa559569": {"__data__": {"id_": "5ef454ee-9d20-4491-ba57-d0d4aa559569", "embedding": null, "metadata": {"page_label": "1", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the article \"Corporate Relative Valuation using Heterogeneous Multi-Modal Graph Neural Network\"?\n2. How does the availability of information about a company's invisible assets enable a new paradigm for learning and evaluating corporate relative values automatically?\n3. What is the methodology used in the article to develop a heterogeneous multi-modal graph neural network for corporate relative valuation?", "section_summary": "The section discusses the article \"Corporate Relative Valuation using Heterogeneous Multi-Modal Graph Neural Network\" by Yang et al. The purpose of the article is to develop a new paradigm for learning and evaluating corporate relative values automatically using the availability of information about a company's invisible assets, such as patents, talent, and investors. The methodology used in the article involves forming the companies and their core members as a heterogeneous graph with semantically-rich multi-modal data, and developing a heterogeneous multi-modal graph neural network method, named HM2, which deals with embedding challenges involving modal attribute encoding, multi-modal aggregation, and valuation prediction modules. The article highlights the effectiveness of the HM2 method in extracting latent embeddings that reflect domain experts' behavior and are effective for corporate relative valuation.", "excerpt_keywords": "1. Corporate relative valuation, 2. Heterogeneous graph neural network, 3. Multi-modal data, 4. Invisible assets, 5. Patents, talent, investors, 6. Domain experts, 7. Network embeddings, 8. Valuation prediction, 9. Non-publicly listed companies, 10. Venture capital firms."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "f40bb129-03c8-42b9-80ee-694879b4e4ae", "node_type": "4", "metadata": {"page_label": "1", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "11c8a3d01644ae0e8e8d6ff334d5f6546809c9553f532379a424055b9d2180e2"}, "3": {"node_id": "316f1e3e-adc4-4999-8696-fc25f0e20b50", "node_type": "1", "metadata": {"page_label": "1", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "3a549764bd354d9c48b3f179a3b70a8dae56274e0c61b849b03d46b547d12f87"}}, "hash": "afeb43fac3703a86579f4bf153211af428a6a322710dec9d1f8f87aa6305abf3", "text": "1041-4347 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TKDE.2021.3080293, IEEE\nTransactions on Knowledge and Data Engineering\n1\nCorporate Relative Valuation using\nHeterogeneous Multi-Modal\nGraph Neural Network\nY ang Y ang, Jia-Qi Y ang, Ran Bao, De-Chuan Zhan, Hengshu Zhu Senior Member, IEEE , Xiao-Ru\nGao, Hui Xiong, Fellow, IEEE and Jian Y ang Member, IEEE\nAbstract\u2014Corporate relative valuation (CRV) refers to the process of comparing a company\u2019s value from company products, core staff\nand other related information, so that we can assess the company\u2019s market value, which is critical for venture capital \ufb01rms. Traditional\nrelative valuation methods heavily rely on tedious and expensive human efforts, especially for non-publicly listed companies. However,\nthe availability of information about company\u2019s invisible assets, such as patents, talent, and investors, enables a new paradigm to learn\nand evaluate corporate relative values automatically. Indeed, in this paper, we reveal that, the companies and their core members can\nnatually be formed as a heterogeneous graph and the attributes of different nodes include semantically-rich multi-modal data, thereby\nwe are able to extract a latent embedding for each company. The network embeddings can re\ufb02ect domain experts\u2019 behavior and are\neffective for corporate relative valuation. Along this line, we develop a heterogeneous multi-modal graph neural network method,\nnamed HM2, which deals with embedding challenges involving modal attribute encoding, multi-modal aggregation, and valuation\nprediction modules. Speci\ufb01cally,", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "316f1e3e-adc4-4999-8696-fc25f0e20b50": {"__data__": {"id_": "316f1e3e-adc4-4999-8696-fc25f0e20b50", "embedding": null, "metadata": {"page_label": "1", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the proposed method for corporate relative valuation using heterogeneous multi-modal graph neural network?\n2. How does the proposed method address the challenges involved in embedding, multi-modal aggregation, and valuation prediction?\n3. What is the effectiveness of the proposed method in improving the performance of corporate relative valuation?", "prev_section_summary": "The section discusses the article \"Corporate Relative Valuation using Heterogeneous Multi-Modal Graph Neural Network\" by Yang et al. The purpose of the article is to develop a new paradigm for learning and evaluating corporate relative values automatically using the availability of information about a company's invisible assets, such as patents, talent, and investors. The methodology used in the article involves forming the companies and their core members as a heterogeneous graph with semantically-rich multi-modal data, and developing a heterogeneous multi-modal graph neural network method, named HM2, which deals with embedding challenges involving modal attribute encoding, multi-modal aggregation, and valuation prediction modules. The article highlights the effectiveness of the HM2 method in extracting latent embeddings that reflect domain experts' behavior and are effective for corporate relative valuation.", "section_summary": "The section discusses a proposed method for corporate relative valuation using a heterogeneous multi-modal graph neural network (HM2). The method addresses challenges in embedding, multi-modal aggregation, and valuation prediction by performing representation learning for heterogeneous neighbors, adopting self-attention for final prediction, and employing dynamic triplet loss with embeddings of competitors as a constraint. The effectiveness of the proposed method is demonstrated through extensive experiments on real-world data. Key entities mentioned include corporate relative valuation, heterogeneous graph, multi-modal learning, linkage-aware, and dynamic triplet loss.", "excerpt_keywords": "1. Corporate Relative Valuation\n2. Heterogeneous Graph\n3. Multi-Modal Learning\n4. Linkage-Aware\n5. Self-Attention Network\n6. Dynamic Triplet Loss\n7. Embedding Challenges\n8. Valuation Prediction Modules\n9. Domain Experts\n10. Enterprise Applications"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "f40bb129-03c8-42b9-80ee-694879b4e4ae", "node_type": "4", "metadata": {"page_label": "1", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "11c8a3d01644ae0e8e8d6ff334d5f6546809c9553f532379a424055b9d2180e2"}, "2": {"node_id": "5ef454ee-9d20-4491-ba57-d0d4aa559569", "node_type": "1", "metadata": {"page_label": "1", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "afeb43fac3703a86579f4bf153211af428a6a322710dec9d1f8f87aa6305abf3"}, "3": {"node_id": "5c3e77ea-e060-4453-861a-af0a9c89a785", "node_type": "1", "metadata": {"page_label": "1", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "5ad45c87f8f1e75d51a73ce03b2f728b5c1a24247d47197612a3bafbd518d714"}}, "hash": "3a549764bd354d9c48b3f179a3b70a8dae56274e0c61b849b03d46b547d12f87", "text": "and their core members can\nnatually be formed as a heterogeneous graph and the attributes of different nodes include semantically-rich multi-modal data, thereby\nwe are able to extract a latent embedding for each company. The network embeddings can re\ufb02ect domain experts\u2019 behavior and are\neffective for corporate relative valuation. Along this line, we develop a heterogeneous multi-modal graph neural network method,\nnamed HM2, which deals with embedding challenges involving modal attribute encoding, multi-modal aggregation, and valuation\nprediction modules. Speci\ufb01cally, HM2\ufb01rstly performs the representation learning for heterogeneous neighbors of the input company by\ntaking relationships among nodes into consideration, which aggregates node attributes via linkage-aware multi-head attention\nmechanism, rather than multi-instance based methods. Then, HM2adopts the self-attention network to aggregate different modal\nembeddings for \ufb01nal prediction, and employs dynamic triplet loss with embeddings of competitors as the constraint. As a result, HM2\ncan explore companies\u2019 intrinsic properties to improve the CRV performance. Extensive experiments on real-world data demonstrate\nthe effectiveness of the proposed HM2.\nIndex Terms\u2014Corporate Relative Valuation, Heterogeneous Graph, Multi-Modal Learning, Linkage-Aware\n!\n1 I NTRODUCTION\nRecent years, we have witnessed the increasing popular-\nity of applying machine learning models in software as a\nservice (SAAS) and various enterprise applications, which\ngreatly reduces the manual cost and improves the operat-\ning ef\ufb01ciency. For example, [1] proposed an intelligent job\ninterview system, which can be applied in human resources\nmanagement (HRM); [2] utilized the structure-aware con-\nvolution neural network for talent \ufb02ow forecast, which can\nbe introduced into enterprise resource planning (ERP); [3]\n\u2022Yang Yang and Jian Yang are with the Nanjing University of Science and\nTechnology, Nanjing 210094,", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "5c3e77ea-e060-4453-861a-af0a9c89a785": {"__data__": {"id_": "5c3e77ea-e060-4453-861a-af0a9c89a785", "embedding": null, "metadata": {"page_label": "1", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What are some specific applications of heterogeneous multi-modal graph neural networks in enterprise management?\n2. How have neural networks been used in human resources management, enterprise resource planning, customer relationship management, and talent flow forecasting?\n3. What are some examples of enterprise service companies that have been developed based on graph neural networks?", "prev_section_summary": "The section discusses a proposed method for corporate relative valuation using a heterogeneous multi-modal graph neural network (HM2). The method addresses challenges in embedding, multi-modal aggregation, and valuation prediction by performing representation learning for heterogeneous neighbors, adopting self-attention for final prediction, and employing dynamic triplet loss with embeddings of competitors as a constraint. The effectiveness of the proposed method is demonstrated through extensive experiments on real-world data. Key entities mentioned include corporate relative valuation, heterogeneous graph, multi-modal learning, linkage-aware, and dynamic triplet loss.", "section_summary": "The section discusses the applications of heterogeneous multi-modal graph neural networks in enterprise management. It highlights specific examples of how neural networks have been used in human resources management, enterprise resource planning, customer relationship management, and talent flow forecasting. The section also mentions examples of enterprise service companies that have been developed based on graph neural networks. The authors of the section are Yang et al., who are affiliated with Nanjing University of Science and Technology and Baidu Talent Intelligence Center. The section is titled \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\" and is published in the journal \"Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf\".", "excerpt_keywords": "1. AI, 2. Machine Learning, 3. Neural Networks, 4. Human Resources Management, 5. Enterprise Resource Planning, 6. Customer Relationship Management, 7. Talent Flow Forecast, 8. Job Interview System, 9. User Recommendation, 10. Intelligent Service"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "f40bb129-03c8-42b9-80ee-694879b4e4ae", "node_type": "4", "metadata": {"page_label": "1", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "11c8a3d01644ae0e8e8d6ff334d5f6546809c9553f532379a424055b9d2180e2"}, "2": {"node_id": "316f1e3e-adc4-4999-8696-fc25f0e20b50", "node_type": "1", "metadata": {"page_label": "1", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "3a549764bd354d9c48b3f179a3b70a8dae56274e0c61b849b03d46b547d12f87"}, "3": {"node_id": "cd77e6c2-e980-4dce-9375-a31aa31ccde4", "node_type": "1", "metadata": {"page_label": "1", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "4b819f9d06a019f7a1599c839290e7c82d0b08c8049021be1e3afbe35436dc63"}}, "hash": "5ad45c87f8f1e75d51a73ce03b2f728b5c1a24247d47197612a3bafbd518d714", "text": "as a\nservice (SAAS) and various enterprise applications, which\ngreatly reduces the manual cost and improves the operat-\ning ef\ufb01ciency. For example, [1] proposed an intelligent job\ninterview system, which can be applied in human resources\nmanagement (HRM); [2] utilized the structure-aware con-\nvolution neural network for talent \ufb02ow forecast, which can\nbe introduced into enterprise resource planning (ERP); [3]\n\u2022Yang Yang and Jian Yang are with the Nanjing University of Science and\nTechnology, Nanjing 210094, China.\nE-mail: yyang,csjyang@njust.edu.cn\n\u2022Jia-Qi Yang, Ran Bao and De-Chuan Zhan are with the Nanjing Univer-\nsity, Nanjing 210023, China.\nE-mail: yangjq@lamda.nju.edu.cn, zhandc@nju.edu.cn, baorana@163.com\n\u2022Hengshu Zhu is with Baidu Talent Intelligence Center, Baidu Inc, Beijing\n100000, China.\nE-mail:zhuhengshu@baidu.com\n\u2022Xiao-Ru Gao and Hui Xiong is with the Management Science and\nInformation Systems Department, Rutgers Business School, Rutgers Uni-\nversity, Newark, NJ 07102, USA.\nE-mail: xg89@business.rutgers.edu, hxiong@rutgers.edu\nYang Yang and Jian Yang are with PCA Lab, Key Lab of Intelligent Perception\nand Systems for High-Dimensional Information of Ministry of Education, and\nJiangsu Key Lab of Image and Video Understanding for Social Security, School\nof Computer Science and Engineering, Nanjing University of Science and\nTechnology. De-Chuan Zhan is the corresponding author.applied neural networks for user recommendation, which\ncan be practiced into customer relationship management\n(CRM), etc. Meanwhile, there also spring up many en-\nterprise service companies based on", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "cd77e6c2-e980-4dce-9375-a31aa31ccde4": {"__data__": {"id_": "cd77e6c2-e980-4dce-9375-a31aa31ccde4", "embedding": null, "metadata": {"page_label": "1", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What are some common methods for corporate valuation and their limitations?\n2. How can heterogeneous multi-modal graph neural networks be used for corporate relative valuation?\n3. What are some enterprise service companies that use artificial intelligence technologies for various purposes?", "prev_section_summary": "The section discusses the applications of heterogeneous multi-modal graph neural networks in enterprise management. It highlights specific examples of how neural networks have been used in human resources management, enterprise resource planning, customer relationship management, and talent flow forecasting. The section also mentions examples of enterprise service companies that have been developed based on graph neural networks. The authors of the section are Yang et al., who are affiliated with Nanjing University of Science and Technology and Baidu Talent Intelligence Center. The section is titled \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\" and is published in the journal \"Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf\".", "section_summary": "The section discusses the importance of corporate valuation in SAAS and the limitations of common methods such as discounted cash flow, economic value added, real options, and price-to-sales. It also introduces heterogeneous multi-modal graph neural networks as a potential solution for corporate relative valuation. The section mentions enterprise service companies such as UiPath and Pymetrics that use artificial intelligence technologies for various purposes.", "excerpt_keywords": "1. Neural networks, 2. User recommendation, 3. Customer relationship management, 4. Artificial intelligence, 5. Enterprise service companies, 6. Data mining, 7. Corporate valuation, 8. Discounted cash flow method, 9. Economic value added method, 10. Real options method"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "f40bb129-03c8-42b9-80ee-694879b4e4ae", "node_type": "4", "metadata": {"page_label": "1", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "11c8a3d01644ae0e8e8d6ff334d5f6546809c9553f532379a424055b9d2180e2"}, "2": {"node_id": "5c3e77ea-e060-4453-861a-af0a9c89a785", "node_type": "1", "metadata": {"page_label": "1", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "5ad45c87f8f1e75d51a73ce03b2f728b5c1a24247d47197612a3bafbd518d714"}, "3": {"node_id": "a11e5164-0951-41fc-9479-c9cccb09e5d8", "node_type": "1", "metadata": {"page_label": "1", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "4f39a624741dbc098f2a19a86b333808c97b4c818ca58025855bb253becc9eec"}}, "hash": "4b819f9d06a019f7a1599c839290e7c82d0b08c8049021be1e3afbe35436dc63", "text": "hxiong@rutgers.edu\nYang Yang and Jian Yang are with PCA Lab, Key Lab of Intelligent Perception\nand Systems for High-Dimensional Information of Ministry of Education, and\nJiangsu Key Lab of Image and Video Understanding for Social Security, School\nof Computer Science and Engineering, Nanjing University of Science and\nTechnology. De-Chuan Zhan is the corresponding author.applied neural networks for user recommendation, which\ncan be practiced into customer relationship management\n(CRM), etc. Meanwhile, there also spring up many en-\nterprise service companies based on arti\ufb01cial intelligence\ntechnologies, for example, UiPath1delivers data mining\ntechniques for document management, contact center, hu-\nman resources, supply chains, etc.; Pymetrics2combines\narti\ufb01cial intelligence technology for intelligent recruitment,\ntalent matching, etc. On the other hand, corporate valuation\nplays an important role in SAAS, which is to evaluate the\nrelative value of companies, and establishes a critical basis\nfor various pricing transactions in enterprise applications.\nThere exist several sophisticated corporate absolute val-\nuation methods, for example, discounted cash \ufb02ow method\n(DCF) [4], economic value added method (EVA) [5], real op-\ntions method (ROA) [6] and price-to-sales method (PS) [7].\nWhile these methods always require historical \ufb01nancial\nstatements of the company, which are dif\ufb01cult to acquire,\nespecially for non-publicly listed companies. On the other\nhand, some other corporate relative valuation methods are\nadopted. These methods usually rely on professionals to\ncomprehensively consider the core resources, members, and\ncompetitors of the company, and then carry on the \ufb01nal\nvaluation. Note that this type of methods can be used to\nestimate the company\u2019s value level without detailed \ufb01nan-\n1. http://www.uipath.com\n2.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "a11e5164-0951-41fc-9479-c9cccb09e5d8": {"__data__": {"id_": "a11e5164-0951-41fc-9479-c9cccb09e5d8", "embedding": null, "metadata": {"page_label": "1", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What are the challenges in acquiring data for corporate relative valuation, especially for non-publicly listed companies?\n2. What are some of the methods used for corporate relative valuation, and how do they differ from each other?\n3. How can professionals use their expertise to carry out a comprehensive valuation of a company's core resources, members, and competitors?", "prev_section_summary": "The section discusses the importance of corporate valuation in SAAS and the limitations of common methods such as discounted cash flow, economic value added, real options, and price-to-sales. It also introduces heterogeneous multi-modal graph neural networks as a potential solution for corporate relative valuation. The section mentions enterprise service companies such as UiPath and Pymetrics that use artificial intelligence technologies for various purposes.", "section_summary": "The section discusses the challenges in acquiring data for corporate relative valuation, especially for non-publicly listed companies. It also explains some of the methods used for corporate relative valuation and how professionals can use their expertise to carry out a comprehensive valuation of a company's core resources, members, and competitors. The section mentions the use of heterogeneous multi-modal graph neural networks for corporate relative valuation.", "excerpt_keywords": "1. Corporate valuation,\n2. Non-publicly listed companies,\n3. Core resources, members, and competitors,\n4. Utility,\n5. Fairness,\n6. Positivity,\n7. Professional valuation,\n8. Financial analysis,\n9. Business strategy,\n10. Company growth."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "f40bb129-03c8-42b9-80ee-694879b4e4ae", "node_type": "4", "metadata": {"page_label": "1", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "11c8a3d01644ae0e8e8d6ff334d5f6546809c9553f532379a424055b9d2180e2"}, "2": {"node_id": "cd77e6c2-e980-4dce-9375-a31aa31ccde4", "node_type": "1", "metadata": {"page_label": "1", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "4b819f9d06a019f7a1599c839290e7c82d0b08c8049021be1e3afbe35436dc63"}}, "hash": "4f39a624741dbc098f2a19a86b333808c97b4c818ca58025855bb253becc9eec", "text": "of the company, which are dif\ufb01cult to acquire,\nespecially for non-publicly listed companies. On the other\nhand, some other corporate relative valuation methods are\nadopted. These methods usually rely on professionals to\ncomprehensively consider the core resources, members, and\ncompetitors of the company, and then carry on the \ufb01nal\nvaluation. Note that this type of methods can be used to\nestimate the company\u2019s value level without detailed \ufb01nan-\n1. http://www.uipath.com\n2. https://www.pymetrics.ai/\nAuthorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 20,2021 at 08:36:01 UTC from IEEE Xplore.  Restrictions apply.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "175214df-164e-4f15-baa4-95cce9ade065": {"__data__": {"id_": "175214df-164e-4f15-baa4-95cce9ade065", "embedding": null, "metadata": {"page_label": "2", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using a heterogeneous multi-modal graph neural network for corporate relative valuation?\n2. How does the use of cartoon characters in the example of company structure help protect privacy?\n3. What are the three factors that are considered in manual relative valuation without financial statements?", "prev_section_summary": "The section discusses the challenges in acquiring data for corporate relative valuation, especially for non-publicly listed companies. It also explains some of the methods used for corporate relative valuation and how professionals can use their expertise to carry out a comprehensive valuation of a company's core resources, members, and competitors. The section mentions the use of heterogeneous multi-modal graph neural networks for corporate relative valuation.", "section_summary": "The section discusses the use of a heterogeneous multi-modal graph neural network for corporate relative valuation (CRV). CRV is a method used to estimate the value of a company based on its structure and other relevant information. The section explains that CRV usually considers two aspects of a company's structure: affiliates and members. To protect privacy, the section uses cartoon characters to replace the companies and members. The section also discusses the limitations of manual CRV without financial statements and the need for automatic or semi-automatic CRV technology using machine learning models. The section mentions that manual CRV without financial statements considers three factors: the core resources of the company and its affiliates, the business conditions, and the competitive landscape.", "excerpt_keywords": "1. CRV, 2. company structure, 3. affiliate information, 4. member information, 5. cartoon characters, 6. heterogeneous graph, 7. statement analysis, 8. venture capital, 9. company valuation, 10. machine learning models."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "7d6b85a8-1c1c-4b4a-8c9f-9827ff561d18", "node_type": "4", "metadata": {"page_label": "2", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "39233ab9964ec63ed25c7b38225e24fa8dd469d5b6424a91d03b0aeb40f84b69"}, "3": {"node_id": "f33ded4d-b4d3-4685-93b3-bc7231ac85e8", "node_type": "1", "metadata": {"page_label": "2", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "dd4c81bdf5cc720a0182dfa50b3ac880c6e0a70d9a2cba9d3a5b9796c4868082"}}, "hash": "3bfb07c09e85dbd9a8598da2b1a4c303d2f4614ecf9e4da33ebcc4487bc7307e", "text": "1041-4347 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TKDE.2021.3080293, IEEE\nTransactions on Knowledge and Data Engineering\n2\nholding\ninvestmentmanager\nsupervisormanager\nsupervisordirector\nsupervisor\ndirectormanager\nmanagersupervisorholding\nFig. 1. (Best viewed in color.) Example of company structure. CRV usu-\nally considers two aspects of the company\u2019s structure: 1) af\ufb01liates, i.e.,\nthe relevant information of the company and its af\ufb01liates (for example,\nfounded, acquired, and invested subsidiaries); 2) members, i.e., the\nrelevant information of company\u2019s core member (for example, manager,\nsupervisor, etc). Considering privacy, we use cartoon characters for re-\nplacement. Note that the companies and members can be regarded as\nentities, and the connections among them can be regarded as linkages,\nthereby all the data can naturally be constructed to a heterogeneous\ngraph.\ncial statement analysis, whereas needs precise judgments\nand heavily relies on tedious and expensive manpower.\nWith the economic development, the number of companies\nhas increased dramatically, thereby it is undoubtedly dif\ufb01-\ncult for venture capital \ufb01rms to conduct large amount of\ncompany valuation screening on interested companies. In\nresult, it urgently needs automatic or semi-automatic CRV\ntechnology by applying machine learning models.\nAs a matter of fact, relative valuation performed man-\nually without \ufb01nancial statements always considers three\nfactors: 1) the core resources of the company and its af\ufb01li-\nates, such as the basic information, business conditions,", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "f33ded4d-b4d3-4685-93b3-bc7231ac85e8": {"__data__": {"id_": "f33ded4d-b4d3-4685-93b3-bc7231ac85e8", "embedding": null, "metadata": {"page_label": "2", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of developing a heterogeneous graph neural network for corporate relative valuation?\n2. How does the heterogeneous multi-modal graph constructed by companies and members aid in obtaining new latent embeddings for corporate valuation?\n3. What are the different types of nodes and graph linkages in the heterogeneous multi-modal graph used for corporate relative valuation?", "prev_section_summary": "The section discusses the use of a heterogeneous multi-modal graph neural network for corporate relative valuation (CRV). CRV is a method used to estimate the value of a company based on its structure and other relevant information. The section explains that CRV usually considers two aspects of a company's structure: affiliates and members. To protect privacy, the section uses cartoon characters to replace the companies and members. The section also discusses the limitations of manual CRV without financial statements and the need for automatic or semi-automatic CRV technology using machine learning models. The section mentions that manual CRV without financial statements considers three factors: the core resources of the company and its affiliates, the business conditions, and the competitive landscape.", "section_summary": "The section discusses the development of a heterogeneous graph neural network (HM2) for corporate relative valuation (CRV). The purpose of developing HM2 is to automate or semi-automate the CRV process, which is currently difficult due to the large number of companies and the need for financial statements. The heterogeneous multi-modal graph constructed by companies and members aids in obtaining new latent embeddings for corporate valuation. The graph consists of two types of nodes: companies and members, with different types of attributes and graph linkages. HM2 is a deep graph network that can acquire discriminative embeddings of the company node by encoding heterogeneous neighbors comprehensively.", "excerpt_keywords": "1. Heterogeneous graph neural network\n2. Corporate relative valuation\n3. Machine learning models\n4. Multi-modal data\n5. Company valuation\n6. Financial statements\n7. Core resources\n8. Core members\n9. Competitors\n10. Latent embeddings"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "7d6b85a8-1c1c-4b4a-8c9f-9827ff561d18", "node_type": "4", "metadata": {"page_label": "2", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "39233ab9964ec63ed25c7b38225e24fa8dd469d5b6424a91d03b0aeb40f84b69"}, "2": {"node_id": "175214df-164e-4f15-baa4-95cce9ade065", "node_type": "1", "metadata": {"page_label": "2", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "3bfb07c09e85dbd9a8598da2b1a4c303d2f4614ecf9e4da33ebcc4487bc7307e"}, "3": {"node_id": "ab91e265-e1cc-463a-9277-8eb7f9086794", "node_type": "1", "metadata": {"page_label": "2", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "0d3f9201045f0061d543df3e2c4453e99a79cf1cd665792ff33bc53d65a1e106"}}, "hash": "dd4c81bdf5cc720a0182dfa50b3ac880c6e0a70d9a2cba9d3a5b9796c4868082", "text": "manpower.\nWith the economic development, the number of companies\nhas increased dramatically, thereby it is undoubtedly dif\ufb01-\ncult for venture capital \ufb01rms to conduct large amount of\ncompany valuation screening on interested companies. In\nresult, it urgently needs automatic or semi-automatic CRV\ntechnology by applying machine learning models.\nAs a matter of fact, relative valuation performed man-\nually without \ufb01nancial statements always considers three\nfactors: 1) the core resources of the company and its af\ufb01li-\nates, such as the basic information, business conditions, and\nintellectual properties; 2) the information of core members\nof the company, such as member\u2019s background, resume, and\nin\ufb02uence; and 3) the valuation of competitors within the\nsame industry. Naturally, as shown in Figure 1, these com-\npanies and members construct a complex heterogeneous\nmulti-modal graph. In detail, there are two types of nodes\nin the graph, i.e., companies and members. Meanwhile, at-\ntributes of nodes constitute multi-modal data, i.e., different\ntypes of nodes have various descriptions. Besides, there\nappear multiple types of graph linkages, i.e., company-\ncompany, company-member, and member-member. There-\nfore, by comprehensively modeling the corporate/personal\nattributes and the linkages among them, we can obtain new\nlatent embeddings to describe the company, which can be\nfurther utilized in corporate valuation task. This learning\nprocedure is also con\ufb01rmed with professional domain ex-\nperts\u2019 operation in reality.\nInspired by the observations above, we develop HM2,\na heterogeneous graph neural network for corporate rela-tive valuation. HM2is a deep graph network, which can\nacquire discriminative embeddings of the company node by\nencoding heterogeneous neighbors comprehensively. Differ-\nent from previous HGNNs, HM2can effectively capture\nthe relationships among nodes, and design", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "ab91e265-e1cc-463a-9277-8eb7f9086794": {"__data__": {"id_": "ab91e265-e1cc-463a-9277-8eb7f9086794", "embedding": null, "metadata": {"page_label": "2", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the main contribution of the paper \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\"?\n2. How does the Heterogeneous Multi-Modal Graph Neural Network (HM2) differ from previous Heterogeneous Graph Neural Networks (HGNNs) in terms of its ability to capture relationships among nodes and improve final performance?\n3. What is the purpose of the extra triplet loss in the loss function of HM2, and how does it enhance the embedding presentation capability of the network?", "prev_section_summary": "The section discusses the development of a heterogeneous graph neural network (HM2) for corporate relative valuation (CRV). The purpose of developing HM2 is to automate or semi-automate the CRV process, which is currently difficult due to the large number of companies and the need for financial statements. The heterogeneous multi-modal graph constructed by companies and members aids in obtaining new latent embeddings for corporate valuation. The graph consists of two types of nodes: companies and members, with different types of attributes and graph linkages. HM2 is a deep graph network that can acquire discriminative embeddings of the company node by encoding heterogeneous neighbors comprehensively.", "section_summary": "The section discusses the paper \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\" by Yang et al. The main contribution of the paper is the development of HM2, a heterogeneous graph neural network for corporate relative valuation. HM2 is a deep graph network that can acquire discriminative embeddings of the company node by encoding heterogeneous neighbors comprehensively. The network utilizes adaptive weighted ensemble to aggregate multi-modal node embedding, which can capture modal interactions and get more descriptive capabilities. The loss function includes the extra triplet loss, which considers the structure with competitors' embeddings except for normal company valuation loss, and aims to enhance the embedding presentation capability by multi-task operator. The paper formalizes the corporate relative valuation as the heterogeneous multi-modal graph structure, which includes heterogeneous nodes, linkages, and multi-modal node attributes in specific.", "excerpt_keywords": "1. Corporate relative valuation, 2. Heterogeneous graph neural network, 3. Multi-head attention mechanism, 4. Adaptive weighted ensemble, 5. Triplet loss, 6. Heterogeneous nodes, 7. Linkages, 8. Multi-modal node attributes, 9. Inductive learning, 10. Professional domain experts."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "7d6b85a8-1c1c-4b4a-8c9f-9827ff561d18", "node_type": "4", "metadata": {"page_label": "2", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "39233ab9964ec63ed25c7b38225e24fa8dd469d5b6424a91d03b0aeb40f84b69"}, "2": {"node_id": "f33ded4d-b4d3-4685-93b3-bc7231ac85e8", "node_type": "1", "metadata": {"page_label": "2", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "dd4c81bdf5cc720a0182dfa50b3ac880c6e0a70d9a2cba9d3a5b9796c4868082"}, "3": {"node_id": "045590d6-758e-4abb-86e8-8a3a02191a5f", "node_type": "1", "metadata": {"page_label": "2", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "6beaa3488ad35ce5022a7874772fe2b5844b639a881ab5a34e405baea798fba7"}}, "hash": "0d3f9201045f0061d543df3e2c4453e99a79cf1cd665792ff33bc53d65a1e106", "text": "company, which can be\nfurther utilized in corporate valuation task. This learning\nprocedure is also con\ufb01rmed with professional domain ex-\nperts\u2019 operation in reality.\nInspired by the observations above, we develop HM2,\na heterogeneous graph neural network for corporate rela-tive valuation. HM2is a deep graph network, which can\nacquire discriminative embeddings of the company node by\nencoding heterogeneous neighbors comprehensively. Differ-\nent from previous HGNNs, HM2can effectively capture\nthe relationships among nodes, and design speci\ufb01c struc-\ntural loss function to improve the \ufb01nal performance. In\ndetail, based on obtained heterogeneous neighbors of the\ninput company, HM2aggregates node attributes via the\nlinkage-aware multi-head attention mechanism [8], which\neffectively incorporates the relationships into node embed-\nding. Then, HM2utilizes adaptive weighted ensemble to\naggregate multi-modal node embedding, which can capture\nmodal interactions and get more descriptive capabilities.\nMoreover, the loss function includes the extra triplet loss,\nwhich considers the structure with competitors\u2019 embed-\ndings except for normal company valuation loss, and aims\nto enhance the embedding presentation capability by multi-\ntask operator. To the best of our knowledge, we are the\n\ufb01rst to formalize the corporate relative valuation into an in-\nductive learning problem considering heterogeneous graph\nstructure. To summarize, the main contributions are:\n\u2022We formalize the corporate relative valuation as the\nheterogeneous multi-modal graph structure, which in-\ncludes heterogeneous nodes, linkages and multi-modal\nnode attributes in speci\ufb01c;\n\u2022We develop HM2, a heterogeneous multi-modal graph\nneural network, which considers heterogeneous nodes\nand linkages for node embeddings comprehensively,\nand combines speci\ufb01c structure loss for \ufb01nal prediction.\nIn result, HM2can be effectively applied to", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "045590d6-758e-4abb-86e8-8a3a02191a5f": {"__data__": {"id_": "045590d6-758e-4abb-86e8-8a3a02191a5f", "embedding": null, "metadata": {"page_label": "2", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the main contribution of the paper \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\"?\n2. How does the heterogeneous multi-modal graph neural network (HM2) consider heterogeneous nodes and linkages for node embeddings comprehensively?\n3. What is the effectiveness of HM2 in predicting corporate relative valuation, as demonstrated through extensive experiments on a real-world corporate valuation dataset?", "prev_section_summary": "The section discusses the paper \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\" by Yang et al. The main contribution of the paper is the development of HM2, a heterogeneous graph neural network for corporate relative valuation. HM2 is a deep graph network that can acquire discriminative embeddings of the company node by encoding heterogeneous neighbors comprehensively. The network utilizes adaptive weighted ensemble to aggregate multi-modal node embedding, which can capture modal interactions and get more descriptive capabilities. The loss function includes the extra triplet loss, which considers the structure with competitors' embeddings except for normal company valuation loss, and aims to enhance the embedding presentation capability by multi-task operator. The paper formalizes the corporate relative valuation as the heterogeneous multi-modal graph structure, which includes heterogeneous nodes, linkages, and multi-modal node attributes in specific.", "section_summary": "The section discusses the paper \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\" by Yang et al. The main contribution of the paper is the development of a heterogeneous multi-modal graph neural network (HM2) that considers heterogeneous nodes and linkages for node embeddings comprehensively and combines specific structure loss for final prediction. The effectiveness of HM2 in predicting corporate relative valuation is demonstrated through extensive experiments on a real-world corporate valuation dataset. The section also introduces the motivation for the paper, the definition of corporate relative valuation with a heterogeneous company-member graph, and the adopted real-world data. Additionally, the section discusses existing heterogeneous graph neural networks.", "excerpt_keywords": "1. Corporate relative valuation\n2. Heterogeneous multi-modal graph\n3. Heterogeneous graph neural networks\n4. Company-member graph\n5. Real-world data\n6. Citation network\n7. Entity-linkage graph\n8. Unstructured data\n9. Complex topological structure\n10. Dynamic graph"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "7d6b85a8-1c1c-4b4a-8c9f-9827ff561d18", "node_type": "4", "metadata": {"page_label": "2", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "39233ab9964ec63ed25c7b38225e24fa8dd469d5b6424a91d03b0aeb40f84b69"}, "2": {"node_id": "ab91e265-e1cc-463a-9277-8eb7f9086794", "node_type": "1", "metadata": {"page_label": "2", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "0d3f9201045f0061d543df3e2c4453e99a79cf1cd665792ff33bc53d65a1e106"}, "3": {"node_id": "c83c8670-2419-4572-9a81-a859ba0021db", "node_type": "1", "metadata": {"page_label": "2", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "e2f401db2e44f023edd75719be93e605b80866e603808a8df952e545c791d2f2"}}, "hash": "6beaa3488ad35ce5022a7874772fe2b5844b639a881ab5a34e405baea798fba7", "text": "graph\nstructure. To summarize, the main contributions are:\n\u2022We formalize the corporate relative valuation as the\nheterogeneous multi-modal graph structure, which in-\ncludes heterogeneous nodes, linkages and multi-modal\nnode attributes in speci\ufb01c;\n\u2022We develop HM2, a heterogeneous multi-modal graph\nneural network, which considers heterogeneous nodes\nand linkages for node embeddings comprehensively,\nand combines speci\ufb01c structure loss for \ufb01nal prediction.\nIn result, HM2can be effectively applied to corporate\nrelative valuation;\n\u2022We conduct extensive experiments on collected real-\nworld corporate valuation dataset, and our results\ndemonstrate the effectiveness of HM2.\n2 P RELIMINARIES\nIn this section, we declare our motivation, deliver the\nde\ufb01nition of CRV with heterogeneous company-member\ngraph, and then introduce the adopted real-world data. In\naddition, we also introduce existing heterogeneous graph\nneural networks.\n2.1 Motivation\nIn real applications, as shown in Figure 1, relative valuation\nperformed manually always considers two factors [9, 10]:\n1) the company and its af\ufb01liates. 2) the core members of\nthe company. Therefore, similar to the citation network\n(author-article) [11, 12], the companies and members can be\nregarded as entities, and the connections among them can\nbe regarded as linkages. Consequently, the data naturally\nconstruct a complex heterogeneous multi-modal graph. Es-\nsentially, it is unstructured data for the reason that: 1) the\ngraph size is arbitrary, the topological structure is complex,\nand there is no spatial locality like images; 2) the graph does\nnot have a \ufb01xed order of nodes; and 3) the graph is dynamic\nand contains multi-modal features. If we directly concate-\nnate the company embedding and the member embed-\nding as a single example, the neighbor representation and\nstructural information of", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "c83c8670-2419-4572-9a81-a859ba0021db": {"__data__": {"id_": "c83c8670-2419-4572-9a81-a859ba0021db", "embedding": null, "metadata": {"page_label": "2", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What are the challenges associated with using traditional graph neural networks for corporate relative valuation analysis?\n2. How does the heterogeneity of the multi-modal graph affect the accuracy of the corporate relative valuation model?\n3. What are the potential benefits of using a heterogeneous multi-modal graph neural network for corporate relative valuation analysis?", "prev_section_summary": "The section discusses the paper \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\" by Yang et al. The main contribution of the paper is the development of a heterogeneous multi-modal graph neural network (HM2) that considers heterogeneous nodes and linkages for node embeddings comprehensively and combines specific structure loss for final prediction. The effectiveness of HM2 in predicting corporate relative valuation is demonstrated through extensive experiments on a real-world corporate valuation dataset. The section also introduces the motivation for the paper, the definition of corporate relative valuation with a heterogeneous company-member graph, and the adopted real-world data. Additionally, the section discusses existing heterogeneous graph neural networks.", "section_summary": "The section discusses the challenges associated with using traditional graph neural networks for corporate relative valuation analysis, the impact of heterogeneity in a multi-modal graph on the accuracy of the model, and the potential benefits of using a heterogeneous multi-modal graph neural network for corporate relative valuation analysis. The section also highlights the complexity and unstructured nature of the data in a heterogeneous multi-modal graph, and the limitations of directly concatenating company and member embeddings.", "excerpt_keywords": "1. Graph,\n2. Multi-modal,\n3. Unstructured,\n4. Complex,\n5. Heterogeneous,\n6. Topological,\n7. Dynamic,\n8. Neighbor representation,\n9. Structural information,\n10. Company embedding."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "7d6b85a8-1c1c-4b4a-8c9f-9827ff561d18", "node_type": "4", "metadata": {"page_label": "2", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "39233ab9964ec63ed25c7b38225e24fa8dd469d5b6424a91d03b0aeb40f84b69"}, "2": {"node_id": "045590d6-758e-4abb-86e8-8a3a02191a5f", "node_type": "1", "metadata": {"page_label": "2", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "6beaa3488ad35ce5022a7874772fe2b5844b639a881ab5a34e405baea798fba7"}}, "hash": "e2f401db2e44f023edd75719be93e605b80866e603808a8df952e545c791d2f2", "text": "Consequently, the data naturally\nconstruct a complex heterogeneous multi-modal graph. Es-\nsentially, it is unstructured data for the reason that: 1) the\ngraph size is arbitrary, the topological structure is complex,\nand there is no spatial locality like images; 2) the graph does\nnot have a \ufb01xed order of nodes; and 3) the graph is dynamic\nand contains multi-modal features. If we directly concate-\nnate the company embedding and the member embed-\nding as a single example, the neighbor representation and\nstructural information of the sample cannot be considered.\nAuthorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 20,2021 at 08:36:01 UTC from IEEE Xplore.  Restrictions apply.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "b3b9bab2-09dd-48d1-aede-2769a88f6bac": {"__data__": {"id_": "b3b9bab2-09dd-48d1-aede-2769a88f6bac", "embedding": null, "metadata": {"page_label": "3", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of developing a deep heterogeneous graph method for Corporate Relative Valuation (CRV)?\n2. How does the heterogeneous multi-modal graph differ from a homogeneous graph in terms of node types and linkage representations?\n3. What are the different types of nodes and linkages in a heterogeneous multi-modal graph, and how are their attributes represented?", "prev_section_summary": "The section discusses the challenges associated with using traditional graph neural networks for corporate relative valuation analysis, the impact of heterogeneity in a multi-modal graph on the accuracy of the model, and the potential benefits of using a heterogeneous multi-modal graph neural network for corporate relative valuation analysis. The section also highlights the complexity and unstructured nature of the data in a heterogeneous multi-modal graph, and the limitations of directly concatenating company and member embeddings.", "section_summary": "The section discusses the development of a deep heterogeneous graph method for Corporate Relative Valuation (CRV) using a heterogeneous multi-modal graph. The heterogeneous multi-modal graph differs from a homogeneous graph in terms of node types and linkage representations, with different types of nodes having various attributes and linkages. The section also defines the heterogeneous multi-modal graph and provides an overview of the problem definition.", "excerpt_keywords": "heterogeneous graph, multi-modal information, content-associated graph, node set, linkage set, node type set, linkage type set, attributes, raw dimensional attribute representations, deep heterogeneous graph method, CRV."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "d543d099-ce2c-4224-b9a4-ad759f438763", "node_type": "4", "metadata": {"page_label": "3", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "caf398a52bcc68240a58e796587532bd0f4a4d87fb673d1c07323be7deb7246c"}, "3": {"node_id": "61fa9f26-9135-48ed-9695-1adab3f9f588", "node_type": "1", "metadata": {"page_label": "3", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "b1d62c37ddfd8039acbbb609d9e1526795117d6ba5b95ac48dce207100f9b164"}}, "hash": "78fdc4e45023779516b9eb7d941fac75b1a6588b24e1d9212fc85bd6ab55a147", "text": "1041-4347 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TKDE.2021.3080293, IEEE\nTransactions on Knowledge and Data Engineering\n3\nFig. 2. (Best viewed in color.) The illustration of heterogeneous multi-\nmodal graph. There exist various types of nodes, i.e., we utilize two types\nhere for simplicity (blue and yellow). Meanwhile, the linkages among\nnodes are also with multiple types, i.e., blue, black, and yellow solid\nlines. Different types of nodes can be represented by various attributes.\nTherefore, in this paper, we develop a deep heterogeneous\ngraph method for CRV . The experimental comparison with\ntraditional methods also veri\ufb01es the effectiveness of the\ngraph embedding.\n2.2 Problem De\ufb01nition\nFirst, we formalize the de\ufb01nition of heterogeneous graph\nwith multi-modal information.\nDe\ufb01nition 1. Heterogeneous Multi-Modal Graph\n(HMMG). (also known as Content-associated\nHeterogeneous Graph [13]) As shown in Figure 2,\nHMMG is de\ufb01ned as a graph G= (V,E,C V,CE)with\nnode setV, linkage set E, node type set CV, and linkage\ntype setCE. Different from homogeneous graph that\nnodes belong to a single type, HMMG owns multiple\nnode types, and different types of nodes have various\nlinkage representations. Moreover, attributes of different\ntypes of nodes constitute multi-modal data, i.e., different\ntypes of nodes have various raw dimensional attribute\nrepresentations.\nWith De\ufb01nition 1, we", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "61fa9f26-9135-48ed-9695-1adab3f9f588": {"__data__": {"id_": "61fa9f26-9135-48ed-9695-1adab3f9f588", "embedding": null, "metadata": {"page_label": "3", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the definition of Heterogeneous Multi-Modal Graph Neural Network (HMMG) and how does it differ from a homogeneous graph?\n2. Can you explain the concept of Corporate Relative Valuation (CRV) and its importance in startup and unlisted companies?\n3. How does HMMG representation learning help in solving the CRV problem without relying on financial statement data?", "prev_section_summary": "The section discusses the development of a deep heterogeneous graph method for Corporate Relative Valuation (CRV) using a heterogeneous multi-modal graph. The heterogeneous multi-modal graph differs from a homogeneous graph in terms of node types and linkage representations, with different types of nodes having various attributes and linkages. The section also defines the heterogeneous multi-modal graph and provides an overview of the problem definition.", "section_summary": "The section discusses the concept of Heterogeneous Multi-Modal Graph Neural Network (HMMG) and its difference from a homogeneous graph. It also explains the concept of Corporate Relative Valuation (CRV) and its importance in startup and unlisted companies. The section highlights how HMMG representation learning helps in solving the CRV problem without relying on financial statement data. The section also provides a summary of the datasets used in the work and the key topics and entities discussed.", "excerpt_keywords": "1. HMMG, 2. Corporate Relative Valuation, 3. CRV, 4. Company, 5. Member, 6. Business Class, 7. ICV, 8. Data Node, 9. Edge, 10. Regression"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "d543d099-ce2c-4224-b9a4-ad759f438763", "node_type": "4", "metadata": {"page_label": "3", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "caf398a52bcc68240a58e796587532bd0f4a4d87fb673d1c07323be7deb7246c"}, "2": {"node_id": "b3b9bab2-09dd-48d1-aede-2769a88f6bac", "node_type": "1", "metadata": {"page_label": "3", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "78fdc4e45023779516b9eb7d941fac75b1a6588b24e1d9212fc85bd6ab55a147"}, "3": {"node_id": "f300981f-69ac-4e0a-b4d7-aa69fe74db59", "node_type": "1", "metadata": {"page_label": "3", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "3a26e7368cb4ff139b0f2464924c0d1503f7eb0947fe8cf82e7e6807139adc74"}}, "hash": "b1d62c37ddfd8039acbbb609d9e1526795117d6ba5b95ac48dce207100f9b164", "text": "is de\ufb01ned as a graph G= (V,E,C V,CE)with\nnode setV, linkage set E, node type set CV, and linkage\ntype setCE. Different from homogeneous graph that\nnodes belong to a single type, HMMG owns multiple\nnode types, and different types of nodes have various\nlinkage representations. Moreover, attributes of different\ntypes of nodes constitute multi-modal data, i.e., different\ntypes of nodes have various raw dimensional attribute\nrepresentations.\nWith De\ufb01nition 1, we can observe that the company\npenetration graph actually is an HMMG with two types of\nnodes. In detail, the node type set CVincludes: company\nand member, and the linkage type set CEincludes: company-\ncompany, company-member and member-member. Then, we can\nde\ufb01ne the corporate relative valuation.\nProblem 1. Corporate Relative Valuation (CRV). CRV aims\nto estimate the relative valuation or value level, i.e.,\na regression or classi\ufb01cation problem, without the \ufb01-\nnancial statement data. CRV is widely used for star-\ntups and unlisted companies, and always considers the\ncore resources, members, and competitors of the input\ncompany. Traditional CRV usually relies on experienced\nexperts.\nIn summary, we now de\ufb01ne the CRV problem with\nHMMG representation learning. Without any loss of gen-\nerality, we provide both approximate (coarse-grained) andTABLE 1\nDatasets used in this work. ICV denotes Internet corporate valuation,\nCRV represents corporate relative valuation level, and BC represents\nbusiness class.\nData Node Edge CRV BC\nICVCompany: 4362\nMember: 6877Company-Company: 5106\nCompany-Member: 13123\nMember-Member: 282244 7\n/uni00000014/uni00000013/uni00000013/uni00000010/uni00000015/uni00000013/uni00000013", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "f300981f-69ac-4e0a-b4d7-aa69fe74db59": {"__data__": {"id_": "f300981f-69ac-4e0a-b4d7-aa69fe74db59", "embedding": null, "metadata": {"page_label": "3", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the study and what are the main findings of the research on Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation?\n2. How does the proposed model compare to other existing methods for corporate relative valuation and what are its advantages and limitations?\n3. What are the potential applications of this model in the field of corporate finance and how can it be used to improve decision-making processes?", "prev_section_summary": "The section discusses the concept of Heterogeneous Multi-Modal Graph Neural Network (HMMG) and its difference from a homogeneous graph. It also explains the concept of Corporate Relative Valuation (CRV) and its importance in startup and unlisted companies. The section highlights how HMMG representation learning helps in solving the CRV problem without relying on financial statement data. The section also provides a summary of the datasets used in the work and the key topics and entities discussed.", "section_summary": "The section discusses a study on the use of Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation. The purpose of the study is to develop a model that can accurately predict the relative valuation of a company based on various factors such as financial statements, news articles, and social media sentiment. The main findings of the research show that the proposed model outperforms existing methods for corporate relative valuation and can provide more accurate predictions. The section also discusses the potential applications of this model in the field of corporate finance and how it can be used to improve decision-making processes.", "excerpt_keywords": "CRV, ICV, business class, corporate relative valuation level, company-company, company-member, member-member, member, edge, node"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "d543d099-ce2c-4224-b9a4-ad759f438763", "node_type": "4", "metadata": {"page_label": "3", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "caf398a52bcc68240a58e796587532bd0f4a4d87fb673d1c07323be7deb7246c"}, "2": {"node_id": "61fa9f26-9135-48ed-9695-1adab3f9f588", "node_type": "1", "metadata": {"page_label": "3", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "b1d62c37ddfd8039acbbb609d9e1526795117d6ba5b95ac48dce207100f9b164"}, "3": {"node_id": "b242dc15-7e7c-4e06-afa3-0a6a68e209c3", "node_type": "1", "metadata": {"page_label": "3", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "23917c953572944b3cff98f9af78278dbc3fde1366271a6f61adf40193f41afb"}}, "hash": "3a26e7368cb4ff139b0f2464924c0d1503f7eb0947fe8cf82e7e6807139adc74", "text": "learning. Without any loss of gen-\nerality, we provide both approximate (coarse-grained) andTABLE 1\nDatasets used in this work. ICV denotes Internet corporate valuation,\nCRV represents corporate relative valuation level, and BC represents\nbusiness class.\nData Node Edge CRV BC\nICVCompany: 4362\nMember: 6877Company-Company: 5106\nCompany-Member: 13123\nMember-Member: 282244 7\n/uni00000014/uni00000013/uni00000013/uni00000010/uni00000015/uni00000013/uni00000013 /uni00000015/uni00000013/uni00000013/uni00000010/uni00000016/uni00000013/uni00000013 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\n(a) CRV\n/uni00000056/uni00000052/uni00000049/uni00000057/uni0000005a/uni00000044/uni00000055/uni00000048", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "b242dc15-7e7c-4e06-afa3-0a6a68e209c3": {"__data__": {"id_": "b242dc15-7e7c-4e06-afa3-0a6a68e209c3", "embedding": null, "metadata": {"page_label": "3", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the study and what problem does it aim to solve?\n2. What is the methodology used in the study and how does it differ from other approaches?\n3. What are the key findings of the study and how do they contribute to the field of corporate relative valuation?", "prev_section_summary": "The section discusses a study on the use of Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation. The purpose of the study is to develop a model that can accurately predict the relative valuation of a company based on various factors such as financial statements, news articles, and social media sentiment. The main findings of the research show that the proposed model outperforms existing methods for corporate relative valuation and can provide more accurate predictions. The section also discusses the potential applications of this model in the field of corporate finance and how it can be used to improve decision-making processes.", "section_summary": "The section discusses a study that uses a heterogeneous multi-modal graph neural network to solve the problem of corporate relative valuation. The study aims to provide accurate and fine-grained valuations of the input. The methodology used in the study involves a graph neural network that takes into account multiple modalities of data, including financial statements, news articles, and social media posts. The key findings of the study contribute to the field of corporate relative valuation by demonstrating the effectiveness of the heterogeneous multi-modal graph neural network approach. The section includes a figure that visualizes the data used in the study, as well as a discussion of the accuracy of the valuations provided by the model.", "excerpt_keywords": "CRV, business category, valuation, accuracy, fine-grained, input, data visualization, relative, number, classes, BC."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "d543d099-ce2c-4224-b9a4-ad759f438763", "node_type": "4", "metadata": {"page_label": "3", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "caf398a52bcc68240a58e796587532bd0f4a4d87fb673d1c07323be7deb7246c"}, "2": {"node_id": "f300981f-69ac-4e0a-b4d7-aa69fe74db59", "node_type": "1", "metadata": {"page_label": "3", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "3a26e7368cb4ff139b0f2464924c0d1503f7eb0947fe8cf82e7e6807139adc74"}, "3": {"node_id": "c010ac8f-8b4b-44ba-8a15-52e55ecbdf4c", "node_type": "1", "metadata": {"page_label": "3", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "b796160b18a64d86ded7257bfdd73e4e4a7e2c6bbfa583f19bfae4d32c0f2b96"}}, "hash": "23917c953572944b3cff98f9af78278dbc3fde1366271a6f61adf40193f41afb", "text": "CRV\n/uni00000056/uni00000052/uni00000049/uni00000057/uni0000005a/uni00000044/uni00000055/uni00000048 /uni00000056/uni00000046/uni0000004c/uni00000048/uni00000051/uni00000057/uni0000004c/uni00000049/uni0000004c/uni00000046\n/uni00000046/uni00000052/uni00000050/uni00000050/uni00000048/uni00000055/uni00000046/uni0000004c/uni00000044/uni0000004f/uni00000048/uni00000010/uni00000055/uni00000048/uni00000057/uni00000044/uni0000004c/uni0000004f/uni0000004c/uni00000051/uni0000004a/uni00000049/uni0000004c/uni00000051/uni00000044/uni00000051/uni00000046/uni0000004c/uni00000044/uni0000004f\n/uni00000048/uni00000051/uni00000057/uni00000048/uni00000055/uni00000057/uni00000044/uni0000004c/uni00000051/uni00000050/uni00000048/uni00000051/uni00000057/uni00000052/uni00000057/uni0000004b/uni00000048/uni00000055/uni00000056/uni00000013/uni00000015/uni00000013/uni00000013/uni00000017/uni00000013/uni00000013/uni00000019/uni00000013/uni00000013/uni0000001b/uni00000013/uni00000013/uni00000014/uni00000013/uni00000013/uni00000013/uni00000014/uni00000015/uni00000013/uni00000013/uni00000014/uni00000017/uni00000013/uni00000013(b) BC\nFig. 3. Data visualization. (a) is the number of each class in relative\nvaluation level (CRV) and (b) is the number of each class in business\ncategory (BC).\naccurate (\ufb01ne-grained) valuations of the input", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "c010ac8f-8b4b-44ba-8a15-52e55ecbdf4c": {"__data__": {"id_": "c010ac8f-8b4b-44ba-8a15-52e55ecbdf4c", "embedding": null, "metadata": {"page_label": "3", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation?\n2. What are the key challenges in designing a model for estimating corporate relative valuation levels using this network?\n3. How does the real-world corporate valuation dataset provided by the business partner contribute to the development of the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation?", "prev_section_summary": "The section discusses a study that uses a heterogeneous multi-modal graph neural network to solve the problem of corporate relative valuation. The study aims to provide accurate and fine-grained valuations of the input. The methodology used in the study involves a graph neural network that takes into account multiple modalities of data, including financial statements, news articles, and social media posts. The key findings of the study contribute to the field of corporate relative valuation by demonstrating the effectiveness of the heterogeneous multi-modal graph neural network approach. The section includes a figure that visualizes the data used in the study, as well as a discussion of the accuracy of the valuations provided by the model.", "section_summary": "The section discusses the purpose, challenges, and contributions of a Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation. The network is designed to estimate the corporate relative valuation level of companies based on their attributes and structural relationships. The real-world corporate valuation dataset provided by a business partner is used to develop the network, and the internet industry is chosen as the domain due to the prevalence of emerging companies and the availability of data. The key challenges in designing the model include learning the company's embedding, which encodes both structural relationships and node attributes.", "excerpt_keywords": "1. Corporate valuation, 2. Multi-modal graph network, 3. Heterogeneous graph, 4. Business category, 5. Relative valuation level, 6. Node attributes, 7. Embedding, 8. Internet industry, 9. Data collection, 10. Relevance island problem"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "d543d099-ce2c-4224-b9a4-ad759f438763", "node_type": "4", "metadata": {"page_label": "3", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "caf398a52bcc68240a58e796587532bd0f4a4d87fb673d1c07323be7deb7246c"}, "2": {"node_id": "b242dc15-7e7c-4e06-afa3-0a6a68e209c3", "node_type": "1", "metadata": {"page_label": "3", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "23917c953572944b3cff98f9af78278dbc3fde1366271a6f61adf40193f41afb"}, "3": {"node_id": "d2f8d11b-631b-4434-8ae4-8da32d3a3437", "node_type": "1", "metadata": {"page_label": "3", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "23824154bc612058ad6f7c0706df44d0735e96262e54eaa9ab55d62511ef7146"}}, "hash": "b796160b18a64d86ded7257bfdd73e4e4a7e2c6bbfa583f19bfae4d32c0f2b96", "text": "BC\nFig. 3. Data visualization. (a) is the number of each class in relative\nvaluation level (CRV) and (b) is the number of each class in business\ncategory (BC).\naccurate (\ufb01ne-grained) valuations of the input companies in\nexperiments.\nDe\ufb01nition 2. Heterogeneous Multi-Modal Graph Network\nfor Corporate Relative Valuation. Given an HMMG\nG= (V,E,C V,CE), each corporate node v1\niinGhas\nits own attribute x1\ni, and is with two ground truth,\ni.e.,yb\ni\u2208RLbdenotes the business category, with Lb\nrepresents the dimension, and yp\ni\u2208RLpdenotes the\ncorporate relative valuation level, with Lpalso denotes\nthe dimension. Besides, each member node v2\njinGalso\nhas corresponding descriptions x2\nj. The task is to design\na modelfthat able to estimate corporate relative valua-\ntion level yp\niof these companies, and the key challenge\noffis to learn company\u2019s embedding, which encodes\nboth structural relationships and node attributes.\nNote that the ambition is to estimate the corporate value,\nthereby we concentrate on the embedding of the company\nnodes, i.e.,v1\ni, in this paper.\n2.3 Data Descriptions\nThe real-world corporate valuation dataset is provided by\nour business partner, and consists of companies in the\ninternet industry. There are several reasons to utilize the\ndata from the internet industry: 1) With the development\nof the Internet, most of the recent emerging companies are\nbelong to the internet industry, and serious data missing\nis a universal problem among companies in other different\ndomains; 2) Considering the cost of data collection, Internet\nindustry takes up the largest number of companies in the\ncollected data; and 3) The heterogeneous graphs of internet\ncompanies have the relevance island problem with the het-\nerogeneous graphs of", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "d2f8d11b-631b-4434-8ae4-8da32d3a3437": {"__data__": {"id_": "d2f8d11b-631b-4434-8ae4-8da32d3a3437", "embedding": null, "metadata": {"page_label": "3", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What are the specific challenges faced by companies in the internet industry when it comes to data collection and analysis?\n2. How does the heterogeneity of graphs in the internet industry differ from those in other domains, and what implications does this have for relative valuation?\n3. What are the potential benefits of using data from the internet industry for corporate relative valuation, and how does this compare to using data from other industries?", "prev_section_summary": "The section discusses the purpose, challenges, and contributions of a Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation. The network is designed to estimate the corporate relative valuation level of companies based on their attributes and structural relationships. The real-world corporate valuation dataset provided by a business partner is used to develop the network, and the internet industry is chosen as the domain due to the prevalence of emerging companies and the availability of data. The key challenges in designing the model include learning the company's embedding, which encodes both structural relationships and node attributes.", "section_summary": "The section discusses the challenges faced by companies in the internet industry when it comes to data collection and analysis, specifically in relation to corporate relative valuation. The heterogeneity of graphs in the internet industry differs from those in other domains, and this has implications for relative valuation. The potential benefits of using data from the internet industry for corporate relative valuation are discussed, including the fact that most emerging companies are in the internet industry, the internet industry takes up the largest number of companies in the collected data, and the heterogeneous graphs of internet companies have relevance island problem with the heterogeneous graphs of other domains. The section also notes that the data is mostly from startups or unlisted companies that are in need of relative valuation.", "excerpt_keywords": "1. Internet industry,\n2. Data collection,\n3. Emerging companies,\n4. Cost of data collection,\n5. Heterogeneous graphs,\n6. Relevance island problem,\n7. Unified graph,\n8. Startups,\n9. Unlisted companies,\n10. Relative valuation."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "d543d099-ce2c-4224-b9a4-ad759f438763", "node_type": "4", "metadata": {"page_label": "3", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "caf398a52bcc68240a58e796587532bd0f4a4d87fb673d1c07323be7deb7246c"}, "2": {"node_id": "c010ac8f-8b4b-44ba-8a15-52e55ecbdf4c", "node_type": "1", "metadata": {"page_label": "3", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "b796160b18a64d86ded7257bfdd73e4e4a7e2c6bbfa583f19bfae4d32c0f2b96"}}, "hash": "23824154bc612058ad6f7c0706df44d0735e96262e54eaa9ab55d62511ef7146", "text": "provided by\nour business partner, and consists of companies in the\ninternet industry. There are several reasons to utilize the\ndata from the internet industry: 1) With the development\nof the Internet, most of the recent emerging companies are\nbelong to the internet industry, and serious data missing\nis a universal problem among companies in other different\ndomains; 2) Considering the cost of data collection, Internet\nindustry takes up the largest number of companies in the\ncollected data; and 3) The heterogeneous graphs of internet\ncompanies have the relevance island problem with the het-\nerogeneous graphs of other domains, i.e., few connections\nbetween two domains\u2019 heterogeneous graphs. Thereby it\nis dif\ufb01cult to consider all domains in one uni\ufb01ed graph.\nNote that these companies are mostly startups or unlisted\ncompanies, and are in need of relative valuation. The data\nAuthorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 20,2021 at 08:36:01 UTC from IEEE Xplore.  Restrictions apply.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "13e77e0e-de3d-4c04-a7fd-9684826b4420": {"__data__": {"id_": "13e77e0e-de3d-4c04-a7fd-9684826b4420", "embedding": null, "metadata": {"page_label": "4", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using a heterogeneous multi-modal graph neural network for corporate relative valuation?\n2. How does the graph representation of the dataset differ from a traditional graph representation, and what are the benefits of using node2vec embeddings for member nodes?\n3. What are the different types of linkages in the graph, and how are they represented?", "prev_section_summary": "The section discusses the challenges faced by companies in the internet industry when it comes to data collection and analysis, specifically in relation to corporate relative valuation. The heterogeneity of graphs in the internet industry differs from those in other domains, and this has implications for relative valuation. The potential benefits of using data from the internet industry for corporate relative valuation are discussed, including the fact that most emerging companies are in the internet industry, the internet industry takes up the largest number of companies in the collected data, and the heterogeneous graphs of internet companies have relevance island problem with the heterogeneous graphs of other domains. The section also notes that the data is mostly from startups or unlisted companies that are in need of relative valuation.", "section_summary": "The section discusses the use of a heterogeneous multi-modal graph neural network for corporate relative valuation. The graph representation of the dataset differs from a traditional graph representation, with node2vec embeddings used for member nodes. The graph consists of two types of nodes: company and core member, with three types of linkages: company-company, company-member, and member-member. The relative valuation level has four categories.", "excerpt_keywords": "HMMG, graph, nodes, attributes, company, core member, linkages, investment ratio, acquisition, one-hot encoding, member-member linkages, relative valuation level, categories, 100 millions, 200 millions, 300 millions, 400 million."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "3290c02d-c27c-44d5-b841-aa705c70db9d", "node_type": "4", "metadata": {"page_label": "4", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "03173012074ee8c8aa5aa5c45c0597ec570c256c427f4f26aedc9cbfb4878045"}, "3": {"node_id": "82aa3f52-44ff-4ff9-8e85-01f0aa4f07dc", "node_type": "1", "metadata": {"page_label": "4", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "9c8495b56d1ce297277487f810f01f6aa85cdf15dd0930cb34fbdf5620e73ac6"}}, "hash": "fdd25a443797eb4aa56631d388462c24ca5f4bd7db993ede83743219e1e73ebf", "text": "1041-4347 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TKDE.2021.3080293, IEEE\nTransactions on Knowledge and Data Engineering\n4\ncan be represented as HMMG in De\ufb01nition 1 naturally. The\noriginal data will be published after permission.\nIn detail, the graph consists two types of nodes: com-\npany and core member, the corporate and member nodes\nare associated with their own attributes, i.e., company has\n132 dimensional features, which cover basic information,\nlegal proceedings, business conditions, intellectual property\nand so on. Member has 5 dimensional features, which are\nextracted from personal information. And member nodes\nare concatenated with 45 dimensional embeddings using\nnode2vec [14]. Besides, the linkages constitutes three types:\n1)company-company linkages have two predicates, i.e., invest-\nment, acquisition. We present the representation of linkage\nas investment ratio, and acquisition is denoted by 1. 2)\ncompany-member linkages have nine predicates, for example,\nchief executive of\ufb01cer (CEO), chief operating of\ufb01cer (COO),\nchief technology of\ufb01cer (CTO), chief \ufb01nancial of\ufb01cer (CFO)\nand related derivative positions. We apply one-hot encoding\nto the company-member linkage. 3) member-member linkages\nhave one predicate, i.e., whether they belong to the same\ncompany. The main statistics of the dataset are shown in\nTable 1.\nMoreover, the relative valuation level has 4 categories,\ni.e., 100 millions to 200 millions, 200 millions to 300 millions,\n300 millions to 400", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "82aa3f52-44ff-4ff9-8e85-01f0aa4f07dc": {"__data__": {"id_": "82aa3f52-44ff-4ff9-8e85-01f0aa4f07dc", "embedding": null, "metadata": {"page_label": "4", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of applying one-hot encoding to the company-member linkage in the dataset?\n2. What is the relevance of instances within each business field in the dataset, and how is it effectively considered?\n3. What is the definition of heterogeneous graph neural networks (HGNN) and how does it differ from traditional graph neural networks?", "prev_section_summary": "The section discusses the use of a heterogeneous multi-modal graph neural network for corporate relative valuation. The graph representation of the dataset differs from a traditional graph representation, with node2vec embeddings used for member nodes. The graph consists of two types of nodes: company and core member, with three types of linkages: company-company, company-member, and member-member. The relative valuation level has four categories.", "section_summary": "The section discusses the application of a heterogeneous graph neural network (HGNN) for corporate relative valuation (CRV) using a dataset containing information about company-member linkages, relative valuation levels, and business categories. The section explains the relevance of instances within each business category and how it is effectively considered. The section also presents a generic definition of HGNN and its key idea, which involves processing different types of neighbors for each input node and aggregating different embeddings into a uniform embedding.", "excerpt_keywords": "1. Heterogeneous Graph Neural Networks (HGNN)\n2. Neighbor Aggregation Architecture\n3. Node Embeddings\n4. Uniform Embeddings\n5. Valuation Categories\n6. Business Categories\n7. One-Hot Encoding\n8. Company-Member Linkages\n9. Member-Member Linkages\n10. LogOperator"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "3290c02d-c27c-44d5-b841-aa705c70db9d", "node_type": "4", "metadata": {"page_label": "4", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "03173012074ee8c8aa5aa5c45c0597ec570c256c427f4f26aedc9cbfb4878045"}, "2": {"node_id": "13e77e0e-de3d-4c04-a7fd-9684826b4420", "node_type": "1", "metadata": {"page_label": "4", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "fdd25a443797eb4aa56631d388462c24ca5f4bd7db993ede83743219e1e73ebf"}, "3": {"node_id": "f8bec8e6-d4da-4ede-96a8-c56cbb71f22f", "node_type": "1", "metadata": {"page_label": "4", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "6a5c81acdb0196981bf6c68715c9251a3407eedc0f222a038d9c7ce6026e8103"}}, "hash": "9c8495b56d1ce297277487f810f01f6aa85cdf15dd0930cb34fbdf5620e73ac6", "text": "operating of\ufb01cer (COO),\nchief technology of\ufb01cer (CTO), chief \ufb01nancial of\ufb01cer (CFO)\nand related derivative positions. We apply one-hot encoding\nto the company-member linkage. 3) member-member linkages\nhave one predicate, i.e., whether they belong to the same\ncompany. The main statistics of the dataset are shown in\nTable 1.\nMoreover, the relative valuation level has 4 categories,\ni.e., 100 millions to 200 millions, 200 millions to 300 millions,\n300 millions to 400 millions, and 400 millions above, and\nthe business class (BC) contains 7 categories, i.e., software\nservice, scienti\ufb01c research and technology service, commer-\ncial service, e-retailing service, \ufb01nancial service, entertain-\nment service, and others. The visualizations of CRV and\nBC are shown in Figure 3, and the \ufb01gure reveals that\nthe instances distribute evenly among valuation categories,\nbut unbalanced among business categories. Therefore, the\nrelevance of instances within each business \ufb01eld needs to be\neffectively considered. Note that we conduct experiments\nwith the real corporate value after using logoperator.\n2.4 Heterogeneous Graph Neural Network\nIn this section, we present a generic de\ufb01nition of het-\nerogeneous graph neural networks (HGNN). HGNN is\nmainly based on neighbor aggregation architecture, which\nemphasizes on processing different types of nodes respec-\ntively [13, 15, 16]. In detail, HGNN usually samples different\ntypes of neighbors for each input node, then encodes them\nrespectively, and \ufb01nally aggregates different embeddings\ninto a uniform embedding. The key idea of HGNN is to\nprocess various types of neighbors for node vi, which can\nbe commonly expressed as [13, 15].\nft(vi)", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "f8bec8e6-d4da-4ede-96a8-c56cbb71f22f": {"__data__": {"id_": "f8bec8e6-d4da-4ede-96a8-c56cbb71f22f", "embedding": null, "metadata": {"page_label": "4", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the main architecture of Heterogeneous Multi-Modal Graph Neural Network (HGNN) and how does it process different types of nodes?\n2. How does the LSTM module in HGNN aggregate different embeddings into a uniform embedding?\n3. What is the role of the forget gate vector, input gate vector, and output gate vector in the LSTM module of HGNN?", "prev_section_summary": "The section discusses the application of a heterogeneous graph neural network (HGNN) for corporate relative valuation (CRV) using a dataset containing information about company-member linkages, relative valuation levels, and business categories. The section explains the relevance of instances within each business category and how it is effectively considered. The section also presents a generic definition of HGNN and its key idea, which involves processing different types of neighbors for each input node and aggregating different embeddings into a uniform embedding.", "section_summary": "The section discusses the Heterogeneous Multi-Modal Graph Neural Network (HGNN) architecture, which is mainly based on neighbor aggregation and emphasizes processing different types of nodes. The LSTM module in HGNN aggregates different embeddings into a uniform embedding by using a forget gate vector, input gate vector, and output gate vector. The section also mentions the Bi-LSTM module used in the LSTM module.", "excerpt_keywords": "Graph neural networks, HGNN, neighbor aggregation architecture, node type, neighbor set, attribute, LSTM, concatenation, forget gate vector, input gate vector, output gate vector, Bi-LSTM."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "3290c02d-c27c-44d5-b841-aa705c70db9d", "node_type": "4", "metadata": {"page_label": "4", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "03173012074ee8c8aa5aa5c45c0597ec570c256c427f4f26aedc9cbfb4878045"}, "2": {"node_id": "82aa3f52-44ff-4ff9-8e85-01f0aa4f07dc", "node_type": "1", "metadata": {"page_label": "4", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "9c8495b56d1ce297277487f810f01f6aa85cdf15dd0930cb34fbdf5620e73ac6"}, "3": {"node_id": "e43d10e5-2e85-4811-bca0-5cbb0dbf0907", "node_type": "1", "metadata": {"page_label": "4", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "85b96d433d7f07c2855943c3f03dcf276425e13143af2ba945dc086dfbe35bf4"}}, "hash": "6a5c81acdb0196981bf6c68715c9251a3407eedc0f222a038d9c7ce6026e8103", "text": "graph neural networks (HGNN). HGNN is\nmainly based on neighbor aggregation architecture, which\nemphasizes on processing different types of nodes respec-\ntively [13, 15, 16]. In detail, HGNN usually samples different\ntypes of neighbors for each input node, then encodes them\nrespectively, and \ufb01nally aggregates different embeddings\ninto a uniform embedding. The key idea of HGNN is to\nprocess various types of neighbors for node vi, which can\nbe commonly expressed as [13, 15].\nft(vi) =1\n|Nt(vi)|\u2211\nv\u2032\u2208Nt(vi)\u2212\u2212\u2212\u2212\u2192LSTM{x(v\u2032)}\u2295\u2190\u2212\u2212\u2212\u2212LSTM{x(v\u2032)},(1)\nwheretdenotes the node type, Nt(\u00b7)is the neighbor set\nof input node,|Nt(\u00b7)| is the set size, x(\u00b7) represents the\nattribute of node, and \u2295denotes concatenation. The single\nLSTM can be formulated as:\nzi=\u03c3(Wzx(v\u2032) +Uzhi\u22121+bz),\ngi=\u03c3(Wgx(v\u2032) +Ughi\u22121+bg),\noi=\u03c3(Wox(v\u2032) +Uohi\u22121+bo),\n\u02c6ci=tanh(W cx(v\u2032) +Uchi\u22121+bc),\nci=gi\u2299ci\u22121+zi\u2299\u02c6ci,\nhi=tanh(c i)\u2299oiwhere hiis the hidden state of i\u2212th node,Wj,Uj,bjj\u2208\n{z,g,o,c}are learnable parameters, and zi,gi,oiare forget\ngate vector, input gate vector, and output gate vector of i\u2212th\nnode respectively. \u2299denoted element-wise product.\nHere LSTM module employs Bi-LSTM [17] to", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "e43d10e5-2e85-4811-bca0-5cbb0dbf0907": {"__data__": {"id_": "e43d10e5-2e85-4811-bca0-5cbb0dbf0907", "embedding": null, "metadata": {"page_label": "4", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using Bi-LSTM in the LSTM module of Heterogeneous Multi-Modal Graph Neural Network (HGNN) for Corporate Relative Valuation (CRV) task?\n2. What are the challenges faced in using Heterogeneous Graph Neural Network (HGNN) for CRV task and how can they be addressed?\n3. How does the neighbor sampling method proposed in the paper address the challenges faced in using HGNN for CRV task?", "prev_section_summary": "The section discusses the Heterogeneous Multi-Modal Graph Neural Network (HGNN) architecture, which is mainly based on neighbor aggregation and emphasizes processing different types of nodes. The LSTM module in HGNN aggregates different embeddings into a uniform embedding by using a forget gate vector, input gate vector, and output gate vector. The section also mentions the Bi-LSTM module used in the LSTM module.", "section_summary": "The section discusses the use of Heterogeneous Multi-Modal Graph Neural Network (HGNN) for Corporate Relative Valuation (CRV) task. The LSTM module in HGNN employs Bi-LSTM to capture deep feature interactions, but it may lose more information during feature embedding due to its multi-instance learning nature. The challenges faced in using HGNN for CRV task include the varying number and type of neighbors for member and company nodes, and the inconsistent linkages among homogeneous and heterogeneous neighbors. The proposed method addresses these challenges by designing an effective neighbor sampling method and corresponding fusion methods.", "excerpt_keywords": "LSTM, Bi-LSTM, HGNN, CRV, heterogeneous graph, node sampling, neighbor sampling, feature fusion, homogeneous neighbors, inconsistent linkages, deep feature interactions, embedding space, mean pooling, multi-instance learning, information loss, comprehensive consideration, effective neighbor sampling, heterogeneous neighbors, modal feature descriptions, inconsistent linkages, deep feature representations, general content embeddings."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "3290c02d-c27c-44d5-b841-aa705c70db9d", "node_type": "4", "metadata": {"page_label": "4", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "03173012074ee8c8aa5aa5c45c0597ec570c256c427f4f26aedc9cbfb4878045"}, "2": {"node_id": "f8bec8e6-d4da-4ede-96a8-c56cbb71f22f", "node_type": "1", "metadata": {"page_label": "4", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "6a5c81acdb0196981bf6c68715c9251a3407eedc0f222a038d9c7ce6026e8103"}, "3": {"node_id": "764a3f60-296f-4bbe-9f8d-611a57d7b14a", "node_type": "1", "metadata": {"page_label": "4", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "faa4ec5c4ca62a39cd2e7e625b5a919bd02c6289102ea00ff5cf060dd073b892"}}, "hash": "85b96d433d7f07c2855943c3f03dcf276425e13143af2ba945dc086dfbe35bf4", "text": "i)\u2299oiwhere hiis the hidden state of i\u2212th node,Wj,Uj,bjj\u2208\n{z,g,o,c}are learnable parameters, and zi,gi,oiare forget\ngate vector, input gate vector, and output gate vector of i\u2212th\nnode respectively. \u2299denoted element-wise product.\nHere LSTM module employs Bi-LSTM [17] to capture\ndeep feature interactions. The Bi-LSTM operates on an un-\nordered content set, which is inspired by previous work [18]\nfor aggregating unordered neighbors. In detail, the LSTM\nbased module \ufb01rst transform different neighbors with the\nsame type into a common embedding space, then employs\nthe Bi-LSTM to accumulate deep feature representations of\nall neighbors, and utilizes a mean pooling operator over\nall hidden states to obtain the general content embeddings.\nHGNN establishes corresponding LSTM models for dif-\nferent types of nodes and fuses them to obtain the \ufb01nal\nfeature embeddings. However, it is notable that Bi-LSTM\nin Equation 1 acts as a multi-instance learning operator,\nwhich only aggregates the information of heterogeneous\nneighbors, yet has not considered the linkages among neigh-\nbors. Therefore, Bi-LSTM may lose more information during\nfeature embedding.\n3 P ROPOSED METHOD\nThe usage of HGNN for CRV task mainly faces the follow-\ning challenges: 1) Nodes in heterogeneous graph connect to\ndifferent types of neighbors, and the number of their neigh-\nbors varies, for example, member nodes usually contains\nmore neighbors than company nodes. Thus, we need to\ndesign an effective neighbor sampling method to consider\nboth the number and type of sampling comprehensively.\n2) Heterogeneous neighbors contain different modal fea-\nture descriptions, and the linkages among homogeneous\nand heterogeneous neighbors are also inconsistent. There-\nfore, we need to design corresponding fusion", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "764a3f60-296f-4bbe-9f8d-611a57d7b14a": {"__data__": {"id_": "764a3f60-296f-4bbe-9f8d-611a57d7b14a", "embedding": null, "metadata": {"page_label": "4", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the main challenge faced by Heterogeneous Multi-Modal Graph Neural Network (HM2) for Corporate Relative Valuation (CRV) task?\n2. How does HM2 address the challenges faced by Heterogeneous Multi-Modal Graph Neural Network (HM2) for Corporate Relative Valuation (CRV) task?\n3. What are the three modules that make up HM2 and what is their purpose?", "prev_section_summary": "The section discusses the use of Heterogeneous Multi-Modal Graph Neural Network (HGNN) for Corporate Relative Valuation (CRV) task. The LSTM module in HGNN employs Bi-LSTM to capture deep feature interactions, but it may lose more information during feature embedding due to its multi-instance learning nature. The challenges faced in using HGNN for CRV task include the varying number and type of neighbors for member and company nodes, and the inconsistent linkages among homogeneous and heterogeneous neighbors. The proposed method addresses these challenges by designing an effective neighbor sampling method and corresponding fusion methods.", "section_summary": "The section discusses the challenges faced by Heterogeneous Multi-Modal Graph Neural Network (HM2) for Corporate Relative Valuation (CRV) task and presents HM2 as a solution. The three modules that make up HM2 are the Modal Attribute Encoding Module, Multi-Modal Aggregation Module, and Valuation Loss Module. The Modal Attribute Encoding Module encodes each type of neighbors separately after neighbor sampling, the Multi-Modal Aggregation Module adaptively aggregates different types of neighbors and input node itself, and the Valuation Loss Module considers the loss of valuation prediction and incorporates the similarities of different corporate node embeddings in the same business category.", "excerpt_keywords": "1. Heterogeneous graph\n2. Neighbor sampling\n3. Modal feature descriptions\n4. Multi-modal aggregation\n5. Valuation loss\n6. Adaptive weights\n7. Node embeddings\n8. Consistency regularization\n9. Business category\n10. Corporate node embeddings"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "3290c02d-c27c-44d5-b841-aa705c70db9d", "node_type": "4", "metadata": {"page_label": "4", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "03173012074ee8c8aa5aa5c45c0597ec570c256c427f4f26aedc9cbfb4878045"}, "2": {"node_id": "e43d10e5-2e85-4811-bca0-5cbb0dbf0907", "node_type": "1", "metadata": {"page_label": "4", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "85b96d433d7f07c2855943c3f03dcf276425e13143af2ba945dc086dfbe35bf4"}}, "hash": "faa4ec5c4ca62a39cd2e7e625b5a919bd02c6289102ea00ff5cf060dd073b892", "text": "of HGNN for CRV task mainly faces the follow-\ning challenges: 1) Nodes in heterogeneous graph connect to\ndifferent types of neighbors, and the number of their neigh-\nbors varies, for example, member nodes usually contains\nmore neighbors than company nodes. Thus, we need to\ndesign an effective neighbor sampling method to consider\nboth the number and type of sampling comprehensively.\n2) Heterogeneous neighbors contain different modal fea-\nture descriptions, and the linkages among homogeneous\nand heterogeneous neighbors are also inconsistent. There-\nfore, we need to design corresponding fusion networks for\nheterogeneous neighbors, and consider the linkages when\nlearning embeddings. 3) Different types of neighbors con-\ntribute differently to node embeddings in heterogeneous\ngraph, thus we need to adaptively learn the weights of\nheterogeneous nodes for \ufb01nal fusion.\nBased on the considerations above, we formally present\nHM2, which consists of there modules: 1) modal attribute\nencoding module, 2) multi-modal aggregation module, and\n3) valuation loss module.\n\u2022Modal Attribute Encoding Module: This module en-\ncodes each type of neighbors respectively after neigh-\nbor sampling, i.e., single modal attribute embedding.\nThe key idea is to take both nodes\u2019 attributes and\ntheir relations into consideration for learning overall\nembedding;\n\u2022Multi-Modal Aggregation Module: This module adap-\ntively aggregates different types of neighbors and input\nnode itself, i.e., the multi-modal embedding weighted\naggregation. The key idea is to learn adaptive weights\nfor each modal information;\n\u2022Valuation Loss Module: This module considers the\nloss of valuation prediction, while incorporates the\nsimilarities of different corporate node embeddings in\nthe same business category, which aims to regularize\nthe consistence.\nAuthorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 20,2021 at 08:36:01 UTC from IEEE Xplore.  Restrictions apply.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "80e29ef9-09f7-4ad0-b123-c0f084375603": {"__data__": {"id_": "80e29ef9-09f7-4ad0-b123-c0f084375603", "embedding": null, "metadata": {"page_label": "5", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the overall architecture of the Heterogeneous Multi-Modal Graph Neural Network (HM2) for corporate relative valuation?\n2. What are the different types of nodes and links in the HM2 model, and how are they encoded and aggregated?\n3. How does the HM2 model develop the loss for corporate valuation with structural triplet regularization?", "prev_section_summary": "The section discusses the challenges faced by Heterogeneous Multi-Modal Graph Neural Network (HM2) for Corporate Relative Valuation (CRV) task and presents HM2 as a solution. The three modules that make up HM2 are the Modal Attribute Encoding Module, Multi-Modal Aggregation Module, and Valuation Loss Module. The Modal Attribute Encoding Module encodes each type of neighbors separately after neighbor sampling, the Multi-Modal Aggregation Module adaptively aggregates different types of neighbors and input node itself, and the Valuation Loss Module considers the loss of valuation prediction and incorporates the similarities of different corporate node embeddings in the same business category.", "section_summary": "The section discusses the Heterogeneous Multi-Modal Graph Neural Network (HM2) for corporate relative valuation. The architecture of HM2 is described, including the different types of nodes and links, and how they are encoded and aggregated. The section also explains how the HM2 model develops the loss for corporate valuation with structural triplet regularization. The symbols used in the section are also defined.", "excerpt_keywords": "1. Heterogeneous Multi-Modal Networks\n2. Multi-Head Attention\n3. Adaptive Attention\n4. Structural Triplet Regularization\n5. Neighbor Sampling\n6. Modal Attribute Encoding\n7. Corporate Valuation\n8. Company-Company Linkage\n9. Company-Member Linkage\n10. Member-Member Linkage"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "f4ace53a-c604-4e78-a848-ab7cd073ebde", "node_type": "4", "metadata": {"page_label": "5", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "25a000acde2316443d608565399296dc71d4ccafe753fc2303632501145e19f2"}, "3": {"node_id": "2f3fed68-fbc2-400e-bda1-721f056a8189", "node_type": "1", "metadata": {"page_label": "5", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a07088d44322f14ac0910b074c6c9419c3008af0c40a13178a3f18b863c61e4c"}}, "hash": "adf66470caf63df2eab7de4e9a7b3b2ec90ad819946edab1bc0162a075b43bfb", "text": "1041-4347 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TKDE.2021.3080293, IEEE\nTransactions on Knowledge and Data Engineering\n5\nX\u2026\n\u2026\n\u2026\nXPrediction LossNeighbor Sampling Modal Attribute Encoding Multi -Modal Aggregation\n\ud835\udefd \u2026\u2026\u2026\\\u2a02\u2026Multi -Head \nAttentionMulti -Head \nAttention\u2026\u2026\n\u2026Multi -Head \nAttentionMulti -Head \nAttention\u2026\nFig. 4. The overall architecture of HM2. From left to right, HM2\ufb01rst samples \ufb01x sized neighbors, which include heterogeneous types. Then it encodes\neach modal attributes via deep network with multi-head attention mechanism, and aggregates multi-modal embedding through adaptive attention.\nFinally, it develops the loss via corporate valuation with structural triplet regularization.\nTABLE 2\nDescription of symbols.\nSym. De\ufb01nition\nV set of nodes with different types, i.e., v1(company),\nv2(member)\nE set of edges with different types\nCVTtypes of node: company and member\nCEMtypes of linkage: company-company, company-member\nmember-member\nx attribute of each node, i.e., x1\u2208Rd1(company),\nx2\u2208Rd2(member)\ny ground truth of each company node, i.e., yb(business\ncategory), yp(valuation level)\nf1 modal attribute encoding module\nql\njthe embedding of j\u2212th node in l\u2212th layer\n\u03b1l,h\ni,jthe learnable weight between nodes", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "2f3fed68-fbc2-400e-bda1-721f056a8189": {"__data__": {"id_": "2f3fed68-fbc2-400e-bda1-721f056a8189", "embedding": null, "metadata": {"page_label": "5", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Multi-Modal Graph Neural Network (HM2) in the context of Corporate Relative Valuation?\n2. How does the HM2 model address the challenges of embedding heterogeneous graphs with multiple types of nodes?\n3. What are the key components of the HM2 model and how do they work together to predict the valuation of a company?", "prev_section_summary": "The section discusses the Heterogeneous Multi-Modal Graph Neural Network (HM2) for corporate relative valuation. The architecture of HM2 is described, including the different types of nodes and links, and how they are encoded and aggregated. The section also explains how the HM2 model develops the loss for corporate valuation with structural triplet regularization. The symbols used in the section are also defined.", "section_summary": "The section discusses the use of a Heterogeneous Multi-Modal Graph Neural Network (HM2) for predicting the valuation of a company in the context of Corporate Relative Valuation. The HM2 model addresses the challenges of embedding heterogeneous graphs with multiple types of nodes by developing separate deep feature learning networks to aggregate information of neighbors, which combines multi-head attention mechanism considering various types of linkages in the learning process. The key components of the HM2 model include a modal attribute encoding module, which aggregates attributes of neighbors, and a multi-modal aggregating module, which comprehensively considers the embeddings of input nodes and their neighbors using self-attention mechanism to acquire final embeddings. The section also provides an overview of the HM2 model and defines the symbols used in the paper.", "excerpt_keywords": "1. Graph Neural Networks (GNNs)\n2. Heterogeneous Graphs\n3. Multi-Modal Aggregation\n4. Attention Mechanism\n5. Deep Learning Networks\n6. Valuation Level\n7. Business Category\n8. Node Embeddings\n9. Predicate of Edge\n10. Self-Attention Mechanism"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "f4ace53a-c604-4e78-a848-ab7cd073ebde", "node_type": "4", "metadata": {"page_label": "5", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "25a000acde2316443d608565399296dc71d4ccafe753fc2303632501145e19f2"}, "2": {"node_id": "80e29ef9-09f7-4ad0-b123-c0f084375603", "node_type": "1", "metadata": {"page_label": "5", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "adf66470caf63df2eab7de4e9a7b3b2ec90ad819946edab1bc0162a075b43bfb"}, "3": {"node_id": "43935afc-b148-4e05-a180-fc64c144d4d7", "node_type": "1", "metadata": {"page_label": "5", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "da4b60a075a09a68edf148106ba274b269806f7055ac719b8a2bf4b608f846c2"}}, "hash": "a07088d44322f14ac0910b074c6c9419c3008af0c40a13178a3f18b863c61e4c", "text": "types\nCVTtypes of node: company and member\nCEMtypes of linkage: company-company, company-member\nmember-member\nx attribute of each node, i.e., x1\u2208Rd1(company),\nx2\u2208Rd2(member)\ny ground truth of each company node, i.e., yb(business\ncategory), yp(valuation level)\nf1 modal attribute encoding module\nql\njthe embedding of j\u2212th node in l\u2212th layer\n\u03b1l,h\ni,jthe learnable weight between nodes iandjinl\u2212th layer\nwithh\u2212th head\nHl number of embedding aggregation head in l\u2212th layer\npi,j embedding of the directed edge between nodes iandj\nati,j predicate of edge between nodes iandj\nf2 multi-modal aggregating module\nAn overview of HM2is shown in Figure 4. Speci\ufb01cally,\nfor the input company node (i.e., dotted blue node), we\nwill \ufb01rst sample heterogeneous neighbors (blue and yel-\nlow nodes) of the node. Secondly, we develop separate\ndeep feature learning networks to aggregate information\nof neighbors, which combines multi-head attention mech-\nanism considering various types of linkages in the learning\nprocess. Finally, we comprehensively consider the embed-\ndings of input nodes and their neighbors using self-attention\nmechanism to acquire \ufb01nal embeddings, which are used for\npredicting the company\u2019s valuation. Table 2 provides the\nde\ufb01nition of symbols used in this paper.\n3.1 Modal Attribute Encoding Module\nThe most critical component of graph neural networks\n(GNNs) [18] is to aggregate attributes of neighbors for\nrepresenting input node. However, heterogeneous graphshave multiple types of nodes, rather than homogeneous\ntype considered in previous methods. Therefore, the em-\nbedding of heterogeneous graph faces two challenges: a)\nsample heterogeneous neighbors for each node", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "43935afc-b148-4e05-a180-fc64c144d4d7": {"__data__": {"id_": "43935afc-b148-4e05-a180-fc64c144d4d7", "embedding": null, "metadata": {"page_label": "5", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the main challenge faced in embedding heterogeneous graphs using graph neural networks (GNNs)?\n2. How does the neighbor construction module in the Heterogeneous Multi-Modal Graph Neural Network (HMMG) address the limitations of first-order neighbors in GNNs?\n3. What is the role of random walk sampling in the neighbor construction module of HMMG for dealing with heterogeneous information?", "prev_section_summary": "The section discusses the use of a Heterogeneous Multi-Modal Graph Neural Network (HM2) for predicting the valuation of a company in the context of Corporate Relative Valuation. The HM2 model addresses the challenges of embedding heterogeneous graphs with multiple types of nodes by developing separate deep feature learning networks to aggregate information of neighbors, which combines multi-head attention mechanism considering various types of linkages in the learning process. The key components of the HM2 model include a modal attribute encoding module, which aggregates attributes of neighbors, and a multi-modal aggregating module, which comprehensively considers the embeddings of input nodes and their neighbors using self-attention mechanism to acquire final embeddings. The section also provides an overview of the HM2 model and defines the symbols used in the paper.", "section_summary": "The section discusses the challenges faced in embedding heterogeneous graphs using graph neural networks (GNNs) and how the Heterogeneous Multi-Modal Graph Neural Network (HMMG) addresses these challenges. The neighbor construction module in HMMG uses random walk sampling to provide more useful structural auxiliary information for input nodes, allowing for more discriminative node embeddings. The section also mentions the limitations of first-order neighbors in GNNs and how HMMG addresses these limitations by sampling heterogeneous neighbors using random walk sampling.", "excerpt_keywords": "1. Graph Neural Networks (GNNs)\n2. Heterogeneous Graphs\n3. Node Embedding\n4. Neighbor Construction\n5. Random Walk Sampling\n6. Modal Attribute Encoding Module\n7. HMMG (Heterogeneous Multi-Modal Graph)\n8. First-order Neighbors\n9. Limited Neighbors\n10. Information Loss"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "f4ace53a-c604-4e78-a848-ab7cd073ebde", "node_type": "4", "metadata": {"page_label": "5", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "25a000acde2316443d608565399296dc71d4ccafe753fc2303632501145e19f2"}, "2": {"node_id": "2f3fed68-fbc2-400e-bda1-721f056a8189", "node_type": "1", "metadata": {"page_label": "5", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a07088d44322f14ac0910b074c6c9419c3008af0c40a13178a3f18b863c61e4c"}, "3": {"node_id": "00a8f7a7-a346-4dc4-b8aa-60cfeeeab5a2", "node_type": "1", "metadata": {"page_label": "5", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "c7b1172c228f54ba6abea690102bc2edd2e309309d7860f12b71bad6243a0718"}}, "hash": "da4b60a075a09a68edf148106ba274b269806f7055ac719b8a2bf4b608f846c2", "text": "embeddings, which are used for\npredicting the company\u2019s valuation. Table 2 provides the\nde\ufb01nition of symbols used in this paper.\n3.1 Modal Attribute Encoding Module\nThe most critical component of graph neural networks\n(GNNs) [18] is to aggregate attributes of neighbors for\nrepresenting input node. However, heterogeneous graphshave multiple types of nodes, rather than homogeneous\ntype considered in previous methods. Therefore, the em-\nbedding of heterogeneous graph faces two challenges: a)\nsample heterogeneous neighbors for each node in HMMG;\nb) construct node encoder for each type of node in HMMG.\nNeighbor Construction\nThe neighbor construction here aims to provide more\nuseful structural auxiliary information for input node, so\nthat learns more discriminative node embedding. The com-\nmon method for neighbor sampling is to sample direct\nneighbors of each node, i.e., \ufb01rst-order neighbors. Neverthe-\nless, \ufb01rst-order neighbors have several limitations as men-\ntioned in [19]: 1) Be susceptible to interference. Nodes have\nlimited \ufb01rst-order neighbors, thus noise neighbors with\nincorrect relations or attributes may impair the embedding;\n2) Information loss. The effects of non-direct neighbors are\nlost by aggregating attributes of direct neighbors only, and\nlimited neighbors may lead to insuf\ufb01cient embedding, for\nexample, node A has \ufb01ve direct neighbors while node B only\nhas three; and 3) Aggregation dif\ufb01culty. Sampling direct\nneighbors is unsuitable for aggregating heterogeneous in-\nformation that contains different modal features. Therefore,\nsampling only direct neighbors may play a negative role.\nHeterogeneous neighbors require different transformations\nto deal with various feature types and dimensions.\nTo solve this problem, inspired from [13, 14], HM2uti-\nlizes the random walk sampling for each node. In detail, it\ncontains two", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "00a8f7a7-a346-4dc4-b8aa-60cfeeeab5a2": {"__data__": {"id_": "00a8f7a7-a346-4dc4-b8aa-60cfeeeab5a2", "embedding": null, "metadata": {"page_label": "5", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using random walk sampling in Heterogeneous Multi-Modal Graph Neural Network (HM2) for corporate relative valuation?\n2. How does HM2 utilize random walk sampling to collect all types of neighbors and group them according to their frequency?\n3. What are the benefits of using HM2 for corporate relative valuation compared to traditional methods that only sample direct neighbors?", "prev_section_summary": "The section discusses the challenges faced in embedding heterogeneous graphs using graph neural networks (GNNs) and how the Heterogeneous Multi-Modal Graph Neural Network (HMMG) addresses these challenges. The neighbor construction module in HMMG uses random walk sampling to provide more useful structural auxiliary information for input nodes, allowing for more discriminative node embeddings. The section also mentions the limitations of first-order neighbors in GNNs and how HMMG addresses these limitations by sampling heterogeneous neighbors using random walk sampling.", "section_summary": "The section discusses the use of Heterogeneous Multi-Modal Graph Neural Network (HM2) for corporate relative valuation. The purpose of using random walk sampling in HM2 for corporate relative valuation is to collect all types of neighbors and group them according to their frequency. The benefits of using HM2 for corporate relative valuation compared to traditional methods that only sample direct neighbors include the ability to aggregate heterogeneous information that contains different modal features and the ability to handle various feature types and dimensions. The section describes the random walk sampling process in detail, which involves sampling a fixed size of neighbors with random walk sampling and grouping different types of neighbors according to their frequency.", "excerpt_keywords": "1. Heterogeneous information\n2. Random walk sampling\n3. Node neighbors\n4. Feature types\n5. Dimensions\n6. Sampling direct neighbors\n7. Aggregation difficulty\n8. Transformations\n9. Company node\n10. Node type"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "f4ace53a-c604-4e78-a848-ab7cd073ebde", "node_type": "4", "metadata": {"page_label": "5", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "25a000acde2316443d608565399296dc71d4ccafe753fc2303632501145e19f2"}, "2": {"node_id": "43935afc-b148-4e05-a180-fc64c144d4d7", "node_type": "1", "metadata": {"page_label": "5", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "da4b60a075a09a68edf148106ba274b269806f7055ac719b8a2bf4b608f846c2"}}, "hash": "c7b1172c228f54ba6abea690102bc2edd2e309309d7860f12b71bad6243a0718", "text": "for\nexample, node A has \ufb01ve direct neighbors while node B only\nhas three; and 3) Aggregation dif\ufb01culty. Sampling direct\nneighbors is unsuitable for aggregating heterogeneous in-\nformation that contains different modal features. Therefore,\nsampling only direct neighbors may play a negative role.\nHeterogeneous neighbors require different transformations\nto deal with various feature types and dimensions.\nTo solve this problem, inspired from [13, 14], HM2uti-\nlizes the random walk sampling for each node. In detail, it\ncontains two steps:\n\u2022Step 1: Sampling \ufb01xed size lof neighbors with random\nwalk sampling. In detail, the sampling process starts\nfrom input node v1\ni\u2208V(superscript 1 represents\nthe company node), and iteratively random walks to\nneighbors of current node or returns to v1\niwith a\nprobability. The process ends until collecting lnodes,\ni.e.,N(v1\ni),|N(v1\ni)|=l;\n\u2022Step 2: Grouping different types of neighbors. For each\ntype, it selects top kt(tis the node type) nodes from\nN(v1\ni)according to frequency.\nWith the procedure above, HM2can collect all types of\nAuthorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 20,2021 at 08:36:01 UTC from IEEE Xplore.  Restrictions apply.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "f2712e0a-b6d7-462c-be38-6574eaf31e7f": {"__data__": {"id_": "f2712e0a-b6d7-462c-be38-6574eaf31e7f", "embedding": null, "metadata": {"page_label": "6", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Multi-Modal Graph Neural Network (HMMG) for Corporate Relative Valuation?\n2. How does the HMMG model handle the heterogeneity of nodes and their attributes in the graph?\n3. What are the benefits of using the HMMG model for corporate relative valuation compared to traditional GNN models?", "prev_section_summary": "The section discusses the use of Heterogeneous Multi-Modal Graph Neural Network (HM2) for corporate relative valuation. The purpose of using random walk sampling in HM2 for corporate relative valuation is to collect all types of neighbors and group them according to their frequency. The benefits of using HM2 for corporate relative valuation compared to traditional methods that only sample direct neighbors include the ability to aggregate heterogeneous information that contains different modal features and the ability to handle various feature types and dimensions. The section describes the random walk sampling process in detail, which involves sampling a fixed size of neighbors with random walk sampling and grouping different types of neighbors according to their frequency.", "section_summary": "The section discusses the use of a Heterogeneous Multi-Modal Graph Neural Network (HMMG) for corporate relative valuation. The purpose of the HMMG model is to handle the heterogeneity of nodes and their attributes in the graph, and to collect high-order member neighbors for each node when constructing neighbors. The model focuses on the embedding learning of corporate nodes, but also takes into account the impact of different types of neighbors on node embeddings. The section also mentions the benefits of using the HMMG model for corporate relative valuation compared to traditional GNN models.", "excerpt_keywords": "1. Graph Neural Networks (GNNs)\n2. Heterogeneous Graphs\n3. Corporate Nodes\n4. Member-Member Linkage\n5. Neighbor Embedding\n6. Modal Attribute Encoding\n7. Multi-Modal Neighbors\n8. Core Resources\n9. Valuation\n10. Deep Graph Neural Networks"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "5b21b249-9b0f-48d8-b500-41fb636aaa28", "node_type": "4", "metadata": {"page_label": "6", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "21cc94a20e40b4e0507178618ddae4bf7e7561cbded605a5704620d292c463da"}, "3": {"node_id": "c61fc7bc-98e3-4656-aa93-f11f9245629b", "node_type": "1", "metadata": {"page_label": "6", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "f5fc81f803cd477dd4f09bbe232bb8c3ff6a499171e089f1af64aeeae42ee649"}}, "hash": "eeb7bc68de0f66b1ee27a66586a65feafcc53c868c167b49f62019f99fbf0c5a", "text": "1041-4347 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TKDE.2021.3080293, IEEE\nTransactions on Knowledge and Data Engineering\n6\nneighbors for each node, and the most frequently visited\nneighbors are selected. Notably, the number of each type of\nnode inN(v1\ni)is constrained, which ensures the balance\nof heterogeneous nodes. Note that HM2focuses on the\nembedding learning of corporate nodes, but the member-\nmember linkage can help to collect high-order member\nneighbors for each node when constructing neighbor.\nModal Attribute Encoding\nThe majority of previous GNN models focus on homo-\ngeneous graphs [20, 21, 22], which ignore the impact of\nnode type. However in HMMG, different types of neigh-\nbors contribute differently to node embeddings. For exam-\nple, mature companies have stronger core resources, thus\nthe attributes of corporate nodes have a greater impact,\nwhereas the core members have a relatively large propor-\ntion of impact in several other companies for valuation.\nOn the other hand, different types of nodes have various\ndimensional attributes, which contain inconsistent physical\nmeanings. Therefore, it is unreasonable to directly aggregate\nheterogeneous neighbors as traditional GNN models. In\nother words, heterogeneous multi-modal neighbors require\ndifferent embedding transformations. To solve this prob-\nlem, [13, 15, 16] attempt to handle heterogeneous graph\nembedding with novel deep graph neural networks, in\nwhich heterogeneous multi-modal neighbors are encoded\nseparately, and aggregated for \ufb01nal embedding. However,\nmost of these methods only encode heterogeneous", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "c61fc7bc-98e3-4656-aa93-f11f9245629b": {"__data__": {"id_": "c61fc7bc-98e3-4656-aa93-f11f9245629b", "embedding": null, "metadata": {"page_label": "6", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the problem with traditional GNN models when dealing with heterogeneous multi-modal graph data?\n2. How does the proposed linkage-aware model f1 address the problem of dealing with heterogeneous multi-modal graph data?\n3. What are the factors considered by the linkage-aware model f1 when aggregating neighboring nodes?", "prev_section_summary": "The section discusses the use of a Heterogeneous Multi-Modal Graph Neural Network (HMMG) for corporate relative valuation. The purpose of the HMMG model is to handle the heterogeneity of nodes and their attributes in the graph, and to collect high-order member neighbors for each node when constructing neighbors. The model focuses on the embedding learning of corporate nodes, but also takes into account the impact of different types of neighbors on node embeddings. The section also mentions the benefits of using the HMMG model for corporate relative valuation compared to traditional GNN models.", "section_summary": "The section discusses the problem of traditional GNN models when dealing with heterogeneous multi-modal graph data, and proposes a linkage-aware model called f1 to address this problem. The linkage-aware model considers two factors: the relationships among nodes and the hierarchical embedding. The model aggregates homogeneous neighboring attributes, considering the linkages among nodes comprehensively. The section also explains how the model works and provides an example of how it can be applied.", "excerpt_keywords": "1. Graph Neural Networks (GNNs)\n2. Heterogeneous graph embedding\n3. Multi-modal neighbors\n4. Linkage-aware model\n5. Attention mechanism\n6. Investment and acquisition relationships\n7. Hierarchical embedding\n8. Direct and non-direct neighbors\n9. Feature propagation process\n10. Corporate and member nodes"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "5b21b249-9b0f-48d8-b500-41fb636aaa28", "node_type": "4", "metadata": {"page_label": "6", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "21cc94a20e40b4e0507178618ddae4bf7e7561cbded605a5704620d292c463da"}, "2": {"node_id": "f2712e0a-b6d7-462c-be38-6574eaf31e7f", "node_type": "1", "metadata": {"page_label": "6", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "eeb7bc68de0f66b1ee27a66586a65feafcc53c868c167b49f62019f99fbf0c5a"}, "3": {"node_id": "011cff06-fd54-4e89-a84c-14a498dfefc7", "node_type": "1", "metadata": {"page_label": "6", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "5a5adaa9f9eb48f34641b3af2cdb89d891e8eac77c4092a79fa7bef9aa9f6076"}}, "hash": "f5fc81f803cd477dd4f09bbe232bb8c3ff6a499171e089f1af64aeeae42ee649", "text": "other hand, different types of nodes have various\ndimensional attributes, which contain inconsistent physical\nmeanings. Therefore, it is unreasonable to directly aggregate\nheterogeneous neighbors as traditional GNN models. In\nother words, heterogeneous multi-modal neighbors require\ndifferent embedding transformations. To solve this prob-\nlem, [13, 15, 16] attempt to handle heterogeneous graph\nembedding with novel deep graph neural networks, in\nwhich heterogeneous multi-modal neighbors are encoded\nseparately, and aggregated for \ufb01nal embedding. However,\nmost of these methods only encode heterogeneous neigh-\nbors with multi-instance based process, without considering\nthe relationships among nodes. But the correlations play an\nimportant role in traditional GNNs, i.e., a weighted metric\nin neighbor aggregation.\nTo model the relationships among neighboring nodes,\ninspired by recent work of attention mechanism [23], we\npropose a linkage-aware model f1, rather than directly em-\nbedding aggregation. Speci\ufb01cally, f1considers two factors:\n1) The relationships among nodes. Different relationships\nplay various roles in embedding, for example, investment\nand acquisition represent different af\ufb01liations between two\ncompanies, and acquisition indicates stronger relation; and\n2) The hierarchical embedding. Direct and non-direct neigh-\nbors have different impacts according to feature propaga-\ntion process mechanism [24], i.e., direct neighboring nodes\nplay relatively more important roles. In summary, f1can\naggregate homogeneous neighboring attributes, considering\nthe linkages among nodes comprehensively. Without any\nloss of generality, different types of nodes can have similar\nmodal attribute encoding modules, i.e., corporate and mem-\nber nodes have similar encoding structures except various\ndimensional input.\nTherefore, as shown in the second part of Figure 4, with\nsampled neighbors, t\u2212th type of neighbors of v1\ni(company)\nare denoted as Nt(v1\ni). We refer to the self-head attention\nmechanism, which performs embedding aggregation", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "011cff06-fd54-4e89-a84c-14a498dfefc7": {"__data__": {"id_": "011cff06-fd54-4e89-a84c-14a498dfefc7", "embedding": null, "metadata": {"page_label": "6", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the self-head attention mechanism in the heterogeneous multi-modal graph neural network for corporate relative valuation?\n2. How does the self-head attention mechanism perform embedding aggregation and attention computation simultaneously?\n3. What is the role of the learnable weight \u03b1l k,j in the self-head attention mechanism?", "prev_section_summary": "The section discusses the problem of traditional GNN models when dealing with heterogeneous multi-modal graph data, and proposes a linkage-aware model called f1 to address this problem. The linkage-aware model considers two factors: the relationships among nodes and the hierarchical embedding. The model aggregates homogeneous neighboring attributes, considering the linkages among nodes comprehensively. The section also explains how the model works and provides an example of how it can be applied.", "section_summary": "The section discusses the use of a heterogeneous multi-modal graph neural network for corporate relative valuation. The self-head attention mechanism is a key component of this network, which performs embedding aggregation and attention computation simultaneously. The learnable weight \u03b1l k,j plays a role in this mechanism, acting as a self-attention operator that computes the attention between nodes k and j. The attention computation is based on the transformed representations in common space from raw attribute, and is a single layer forward neural network.", "excerpt_keywords": "1. Self-head attention, 2. Embedding aggregation, 3. Attention computation, 4. Node index, 5. Hidden layer index, 6. Learnable weight, 7. Raw attribute, 8. Transformed representations, 9. Common space, 10. Neural network."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "5b21b249-9b0f-48d8-b500-41fb636aaa28", "node_type": "4", "metadata": {"page_label": "6", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "21cc94a20e40b4e0507178618ddae4bf7e7561cbded605a5704620d292c463da"}, "2": {"node_id": "c61fc7bc-98e3-4656-aa93-f11f9245629b", "node_type": "1", "metadata": {"page_label": "6", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "f5fc81f803cd477dd4f09bbe232bb8c3ff6a499171e089f1af64aeeae42ee649"}, "3": {"node_id": "2ebd71ad-f2b3-4aec-9f3b-505e92f90d1c", "node_type": "1", "metadata": {"page_label": "6", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "474e62277dbd423d819ea32cad4f364224dead6a3efcd4e99ebe4fd6753fcbce"}}, "hash": "5a5adaa9f9eb48f34641b3af2cdb89d891e8eac77c4092a79fa7bef9aa9f6076", "text": "f1can\naggregate homogeneous neighboring attributes, considering\nthe linkages among nodes comprehensively. Without any\nloss of generality, different types of nodes can have similar\nmodal attribute encoding modules, i.e., corporate and mem-\nber nodes have similar encoding structures except various\ndimensional input.\nTherefore, as shown in the second part of Figure 4, with\nsampled neighbors, t\u2212th type of neighbors of v1\ni(company)\nare denoted as Nt(v1\ni). We refer to the self-head attention\nmechanism, which performs embedding aggregation and\nattention computation simultaneously. Formally, the self-\nhead attention aggregation can be formulated as:\nql\nk=\u2211\nj\u2208Nt(i)\u222a{i}\u03b1l\nk,jql\u22121\nj, (2)\nwherel/kdenotes hidden layer index (l = 1,2,\u00b7\u00b7\u00b7,L) and\nnode index (k\u2208Nt(i)\u222a{i}),\u03b1l\nk,jis a learnable weight\nbetween nodes kandj.ql\u22121\nj denotes the embedding ofnodejofl\u22121-th layer output, where q0\nj= \u03a6(x j)is the\ntransformed representations in common space from raw\nattribute, i.e., q0\u2208Rd.\u03b1l\nk,jacts as self-attention operator,\nwhich is a single layer forward neural network, and can be\nformalized as:\n\u03b1l\nk,j=exp(\n(\u03c9\u22a4\nl(\u03a8(p k,j)\u2225atk,j)[ql\u22121\nk\u2299ql\u22121\nj]))\n\u2211\nn\u2208N t(i)\u222a{i}exp(\n(\u03c9\u22a4\nl(\u03a8(p", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "2ebd71ad-f2b3-4aec-9f3b-505e92f90d1c": {"__data__": {"id_": "2ebd71ad-f2b3-4aec-9f3b-505e92f90d1c", "embedding": null, "metadata": {"page_label": "6", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the mathematical formula for the self-attention operator in a heterogeneous multi-modal graph neural network for corporate relative valuation?\n2. How does the weight matrix and vector point multiplication work in the self-attention operator of a heterogeneous multi-modal graph neural network for corporate relative valuation?\n3. What is the purpose of the one-hot representation of directed edges in a heterogeneous multi-modal graph neural network for corporate relative valuation?", "prev_section_summary": "The section discusses the use of a heterogeneous multi-modal graph neural network for corporate relative valuation. The self-head attention mechanism is a key component of this network, which performs embedding aggregation and attention computation simultaneously. The learnable weight \u03b1l k,j plays a role in this mechanism, acting as a self-attention operator that computes the attention between nodes k and j. The attention computation is based on the transformed representations in common space from raw attribute, and is a single layer forward neural network.", "section_summary": "The section discusses the use of a heterogeneous multi-modal graph neural network for corporate relative valuation. The self-attention operator in this network is described mathematically, including the weight matrix and vector point multiplication. The one-hot representation of directed edges is also explained, as is the purpose of the link predicate and the mapping from raw linkage representation to its embedding. The section also mentions the possibility of extending the architecture to include multiple attribute aggregation heads for improved performance and optimization stability.", "excerpt_keywords": "1. Self-attention operator\n2. Neural network\n3. Weight matrix\n4. Vector point multiplication\n5. One-hot representation\n6. Link predicate\n7. Investment ratio\n8. Company-company linkages\n9. Member-member linkages\n10. Attribute aggregation head"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "5b21b249-9b0f-48d8-b500-41fb636aaa28", "node_type": "4", "metadata": {"page_label": "6", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "21cc94a20e40b4e0507178618ddae4bf7e7561cbded605a5704620d292c463da"}, "2": {"node_id": "011cff06-fd54-4e89-a84c-14a498dfefc7", "node_type": "1", "metadata": {"page_label": "6", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "5a5adaa9f9eb48f34641b3af2cdb89d891e8eac77c4092a79fa7bef9aa9f6076"}, "3": {"node_id": "826655f9-172c-46dc-99bf-1017a8305b63", "node_type": "1", "metadata": {"page_label": "6", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a703b0583c3a3e8dc4bebe3c09efcbb82cf8f1e9f1e9bc8b02da9deabfe782cd"}}, "hash": "474e62277dbd423d819ea32cad4f364224dead6a3efcd4e99ebe4fd6753fcbce", "text": "i.e., q0\u2208Rd.\u03b1l\nk,jacts as self-attention operator,\nwhich is a single layer forward neural network, and can be\nformalized as:\n\u03b1l\nk,j=exp(\n(\u03c9\u22a4\nl(\u03a8(p k,j)\u2225atk,j)[ql\u22121\nk\u2299ql\u22121\nj]))\n\u2211\nn\u2208N t(i)\u222a{i}exp(\n(\u03c9\u22a4\nl(\u03a8(p k,n)\u2225atk,n)[ql\u22121\nk\u2299ql\u22121\nn])),\n(3)\nwhere\u03c9\u22a4\nlrepresents the weight matrix for l\u2212th layer, and\u2299\ndenotes the vector point multiplication. pk,j\u2208Rdpdenotes\nthe one-hot representation of directed edge between nodes\nkandj, and \u03a8(\u00b7) denotes mapping from raw linkage\nrepresentation to its embedding. atk,j\u2208Rrepresents the\nlink predicate, i.e., atk,jis the investment ratio for company-\ncompany linkages, and atk,j= 1 for company-member and\nmember-member linkages. \u2225denotes concatenation opera-\ntor, andpk,j=atk,j=\u03beif there exists no direct linkage\nbetween nodes kandj.\u03beis always with a small value (i.e.,\n10\u22123in experiment).\nWe can also extend f1to a more general architecture, in\nwhich each layer contains a variable number of attribute\naggregation head. And multiple heads can promote the\nperformance and optimization stability. Therefore, Equation\n2 and Equation 3 can be reformulated", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "826655f9-172c-46dc-99bf-1017a8305b63": {"__data__": {"id_": "826655f9-172c-46dc-99bf-1017a8305b63", "embedding": null, "metadata": {"page_label": "6", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. How does the heterogeneous multi-modal graph neural network improve the performance and optimization stability of corporate relative valuation?\n2. What is the role of the learnable weight \u03b1 in the aggregated embedding output of the heterogeneous multi-modal graph neural network?\n3. Can the heterogeneous multi-modal graph neural network measure the relationships between input nodes and different types of neighbors, including direct and non-direct neighbors?", "prev_section_summary": "The section discusses the use of a heterogeneous multi-modal graph neural network for corporate relative valuation. The self-attention operator in this network is described mathematically, including the weight matrix and vector point multiplication. The one-hot representation of directed edges is also explained, as is the purpose of the link predicate and the mapping from raw linkage representation to its embedding. The section also mentions the possibility of extending the architecture to include multiple attribute aggregation heads for improved performance and optimization stability.", "section_summary": "The section discusses the use of a heterogeneous multi-modal graph neural network for corporate relative valuation. The network improves the performance and optimization stability of the task by using a learnable weight \u03b1 in the aggregated embedding output. The network can measure the relationships between input nodes and different types of neighbors, including direct and non-direct neighbors, by extending the f1 function to a more general architecture with multiple attribute aggregation heads. The final aggregated embedding output is computed by performing a weighted aggregation of intermediate embeddings, with the learnable weight \u03b1 effectively overcoming the impact of direct and non-direct neighbors.", "excerpt_keywords": "1. Graph neural networks\n2. Attention mechanisms\n3. Embedding learning\n4. Neighborhood aggregation\n5. Weighted aggregation\n6. Positive reinforcement\n7. Fairness and positivity\n8. Utility-based optimization\n9. Ethical considerations\n10. Company-member and member-member linkages"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "5b21b249-9b0f-48d8-b500-41fb636aaa28", "node_type": "4", "metadata": {"page_label": "6", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "21cc94a20e40b4e0507178618ddae4bf7e7561cbded605a5704620d292c463da"}, "2": {"node_id": "2ebd71ad-f2b3-4aec-9f3b-505e92f90d1c", "node_type": "1", "metadata": {"page_label": "6", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "474e62277dbd423d819ea32cad4f364224dead6a3efcd4e99ebe4fd6753fcbce"}, "3": {"node_id": "3b47d874-dd1a-42d2-a582-34bad3d8b0c9", "node_type": "1", "metadata": {"page_label": "6", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "e8305a66b6aba32cea625d96c0146f0e9b253d2f01e1595b4b23653bddf2d6ce"}}, "hash": "a703b0583c3a3e8dc4bebe3c09efcbb82cf8f1e9f1e9bc8b02da9deabfe782cd", "text": "1 for company-member and\nmember-member linkages. \u2225denotes concatenation opera-\ntor, andpk,j=atk,j=\u03beif there exists no direct linkage\nbetween nodes kandj.\u03beis always with a small value (i.e.,\n10\u22123in experiment).\nWe can also extend f1to a more general architecture, in\nwhich each layer contains a variable number of attribute\naggregation head. And multiple heads can promote the\nperformance and optimization stability. Therefore, Equation\n2 and Equation 3 can be reformulated as:\nql\nk=1\n|Hl|\u2211\nh\u2211\nj\u2208Nt(i)\u222a{i}\u03b1l,h\nk,jql\u22121\nj,\n\u03b1l,h\nk,j=exp(\n(\u03c9\u22a4\nl,h(\u03a8(p k,j)\u2225atk,j)[ql\u22121\nk\u2299ql\u22121\nj]))\n\u2211\nn\u2208N t(i)\u222a{i}exp(\n(\u03c9\u22a4\nl,h(\u03a8(p k,n)\u2225atk,n)[ql\u22121\nk\u2299ql\u22121\nn])),\n(4)\nwherehdenotes the h\u2212th head, and Hlis the number of\nheads inl\u2212th layer. Consequently, we can formalize the \ufb01nal\naggregated embedding output of t\u2212th type of neighboring\nnodes as:\nft\n1(v1\ni) =1\n|Nt(i)|\u2211\nk\u2208Nt(i)qL\nk.(5)\nThus,\u03b1l\nk,jcan well measure the relationships between input\nnodes and different types of neighbors, while considering\nthe impact of direct and non-direct neighbors. f1computes\nthe aggregated embeddings by performing a weighted ag-\ngregation of intermediate, and the learnable weight \u03b1can\neffectively overcome", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "3b47d874-dd1a-42d2-a582-34bad3d8b0c9": {"__data__": {"id_": "3b47d874-dd1a-42d2-a582-34bad3d8b0c9", "embedding": null, "metadata": {"page_label": "6", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Multi-Modal Graph Neural Network in the context of Corporate Relative Valuation?\n2. How does the Multi-Modal Aggregating Module in the Heterogeneous Multi-Modal Graph Neural Network improve the final representation of input nodes?\n3. What are the two problems that the learnable weight \u03b1 in the Heterogeneous Multi-Modal Graph Neural Network helps to overcome?", "prev_section_summary": "The section discusses the use of a heterogeneous multi-modal graph neural network for corporate relative valuation. The network improves the performance and optimization stability of the task by using a learnable weight \u03b1 in the aggregated embedding output. The network can measure the relationships between input nodes and different types of neighbors, including direct and non-direct neighbors, by extending the f1 function to a more general architecture with multiple attribute aggregation heads. The final aggregated embedding output is computed by performing a weighted aggregation of intermediate embeddings, with the learnable weight \u03b1 effectively overcoming the impact of direct and non-direct neighbors.", "section_summary": "The section discusses the use of a Heterogeneous Multi-Modal Graph Neural Network (HMMGNN) for Corporate Relative Valuation. The purpose of the HMMGNN is to improve the final representation of input nodes by aggregating different modal embeddings using a Multi-Modal Aggregating Module. The Multi-Modal Aggregating Module uses adaptive attention to capture more discriminative feature capability and effectively overcome two problems mentioned in the section. The two problems that the learnable weight \u03b1 in the HMMGNN helps to overcome are the impact of direct and non-direct neighbors, and the ability to measure relationships between input nodes and different types of neighbors. The section also provides equations for the embedding output of t\u2212th type of neighboring nodes and the final representation of v1i.", "excerpt_keywords": "1. Neural networks\n2. Embedding\n3. Aggregation\n4. Multi-modal\n5. Attention\n6. Discriminative\n7. Feature\n8. Adaptive\n9. Final representation\n10. Learning"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "5b21b249-9b0f-48d8-b500-41fb636aaa28", "node_type": "4", "metadata": {"page_label": "6", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "21cc94a20e40b4e0507178618ddae4bf7e7561cbded605a5704620d292c463da"}, "2": {"node_id": "826655f9-172c-46dc-99bf-1017a8305b63", "node_type": "1", "metadata": {"page_label": "6", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a703b0583c3a3e8dc4bebe3c09efcbb82cf8f1e9f1e9bc8b02da9deabfe782cd"}}, "hash": "e8305a66b6aba32cea625d96c0146f0e9b253d2f01e1595b4b23653bddf2d6ce", "text": "embedding output of t\u2212th type of neighboring\nnodes as:\nft\n1(v1\ni) =1\n|Nt(i)|\u2211\nk\u2208Nt(i)qL\nk.(5)\nThus,\u03b1l\nk,jcan well measure the relationships between input\nnodes and different types of neighbors, while considering\nthe impact of direct and non-direct neighbors. f1computes\nthe aggregated embeddings by performing a weighted ag-\ngregation of intermediate, and the learnable weight \u03b1can\neffectively overcome the two problems mentioned above.\n3.2 Multi-Modal Aggregating Module\nIn this section, we aim to aggregate different modal em-\nbeddings for \ufb01nal representation. As shown in the third\npart of Figure 4, different from concatenating multi-modal\nembedding directly [25], we turn to design a novel adaptive\nattention based network to capture more discriminative\nfeature capability. Formally, the \ufb01nal representation of v1\ni\ncan be computed as:\nf2(v1\ni) =\u2211\nj\u2208{T,i}\u03b2j\u02c6fj\n1(v1\ni),(6)\nAuthorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 20,2021 at 08:36:01 UTC from IEEE Xplore.  Restrictions apply.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "edd73257-0ab1-45ab-abc8-54d97a7edd0a": {"__data__": {"id_": "edd73257-0ab1-45ab-abc8-54d97a7edd0a", "embedding": null, "metadata": {"page_label": "7", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using adaptive weights in the heterogeneous multi-modal graph neural network for corporate relative valuation?\n2. How is the heterogeneous graph representation structural loss incorporated into the overall loss function for training the HM2 model?\n3. What is the role of the mapping introduced in Section 3.1 in the heterogeneous multi-modal graph neural network for corporate relative valuation?", "prev_section_summary": "The section discusses the use of a Heterogeneous Multi-Modal Graph Neural Network (HMMGNN) for Corporate Relative Valuation. The purpose of the HMMGNN is to improve the final representation of input nodes by aggregating different modal embeddings using a Multi-Modal Aggregating Module. The Multi-Modal Aggregating Module uses adaptive attention to capture more discriminative feature capability and effectively overcome two problems mentioned in the section. The two problems that the learnable weight \u03b1 in the HMMGNN helps to overcome are the impact of direct and non-direct neighbors, and the ability to measure relationships between input nodes and different types of neighbors. The section also provides equations for the embedding output of t\u2212th type of neighboring nodes and the final representation of v1i.", "section_summary": "The section discusses the use of a heterogeneous multi-modal graph neural network (HM2) for corporate relative valuation. The network uses adaptive weights to discover relationships among different modalities, and a heterogeneous graph representation structural loss is incorporated into the overall loss function for training. The mapping introduced in Section 3.1 is also discussed, which is used to concatenate embeddings from different modalities. The section concludes with a multi-task learning approach that combines corporate relative valuation loss and heterogeneous graph representation structural loss for training the HM2 model.", "excerpt_keywords": "1. Adaptive weights, 2. Modal embeddings, 3. Leaky ReLU, 4. Rectified Linear Unit, 5. Parameter, 6. Concatenated embeddings, 7. Heterogeneous graph representation, 8. Multi-task learning, 9. Discriminative representation, 10. Overall loss function."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "b26ce5b7-5cd1-4c4f-ab00-7f0ec4074897", "node_type": "4", "metadata": {"page_label": "7", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "4a573f135fd6e0a296538270df39d45736b84875c1602c0db0139321f7e539ec"}, "3": {"node_id": "d8e60937-b046-4729-83f5-35b2bb6e3498", "node_type": "1", "metadata": {"page_label": "7", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "184e7dc8fbcce3e3a209af53417383f490d23a0694120a070295fb7af96aae2f"}}, "hash": "41fb5363715b09682e8f7e309375712c2befce1b2790cc7c140a1d7d042644c5", "text": "1041-4347 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TKDE.2021.3080293, IEEE\nTransactions on Knowledge and Data Engineering\n7\nwhere\u03b2jis the adaptive weights of each modal embedding,\nwhich aim to discover the relationships among different\nmodalities. \u03b2jcan be formulated as:\n\u03b2j=exp{LeakyReLU (u\u22a4\u02c6fj\n1)}\n\u2211\nkexp{LeakyReLU (uT\u02c6fk\n1)},\nwhere LeakyReLU denotes leaky version of a Recti\ufb01ed\nLinear Unit, and u\u2208R2d\u00d71is the parameter. \u02c6fj\n1\u2208R2d\ndenotes the concatenated embeddings:\n\u02c6fj\n1={\nfj\n1\u2225\u03a6(x1\ni), when j\u0338=i,\n\u03a6(x1\ni)\u2225\u03a6(x1\ni),when j =i.\nwhere \u03a6(x1\ni)is the mapping introduced in Section 3.1, and\n\u2225denotes concatenation operator.\n3.3 Model Training\nTo perform corporate relative valuation for input node, we\ntrain HM2from two aspects: 1) corporate relative valuation\nloss, and 2) heterogeneous graph representation structural\nloss. This constructs a multi-task learning approach, which\ncan learn more discriminative representation. In result, the\noverall loss function", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "d8e60937-b046-4729-83f5-35b2bb6e3498": {"__data__": {"id_": "d8e60937-b046-4729-83f5-35b2bb6e3498", "embedding": null, "metadata": {"page_label": "7", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using a heterogeneous multi-modal graph neural network for corporate relative valuation?\n2. How does the model training process work in this approach?\n3. What is the role of the random walk in graph in generating triplets for the model training process?", "prev_section_summary": "The section discusses the use of a heterogeneous multi-modal graph neural network (HM2) for corporate relative valuation. The network uses adaptive weights to discover relationships among different modalities, and a heterogeneous graph representation structural loss is incorporated into the overall loss function for training. The mapping introduced in Section 3.1 is also discussed, which is used to concatenate embeddings from different modalities. The section concludes with a multi-task learning approach that combines corporate relative valuation loss and heterogeneous graph representation structural loss for training the HM2 model.", "section_summary": "The section discusses the use of a heterogeneous multi-modal graph neural network for corporate relative valuation. The model training process involves two aspects: corporate relative valuation loss and heterogeneous graph representation structural loss. The random walk in graph is used to generate triplets for the model training process. The overall loss function is a multi-task learning approach that can learn more discriminative representation. The section also mentions the distance measurement function used to measure the distance between node embeddings.", "excerpt_keywords": "1. Corporate relative valuation\n2. Heterogeneous graph representation\n3. Multi-task learning\n4. Embedding loss\n5. Distance measurement\n6. Random walk\n7. Triplets\n8. Business category\n9. Domain experts\n10. Embedding structure"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "b26ce5b7-5cd1-4c4f-ab00-7f0ec4074897", "node_type": "4", "metadata": {"page_label": "7", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "4a573f135fd6e0a296538270df39d45736b84875c1602c0db0139321f7e539ec"}, "2": {"node_id": "edd73257-0ab1-45ab-abc8-54d97a7edd0a", "node_type": "1", "metadata": {"page_label": "7", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "41fb5363715b09682e8f7e309375712c2befce1b2790cc7c140a1d7d042644c5"}, "3": {"node_id": "e40e0d0c-2842-4989-9e31-f6cbe20c35ab", "node_type": "1", "metadata": {"page_label": "7", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "e4e94118d0faf899b4763ba56f9bc3dc8acb0139261fa52194a7885f1ac2151f"}}, "hash": "184e7dc8fbcce3e3a209af53417383f490d23a0694120a070295fb7af96aae2f", "text": "j =i.\nwhere \u03a6(x1\ni)is the mapping introduced in Section 3.1, and\n\u2225denotes concatenation operator.\n3.3 Model Training\nTo perform corporate relative valuation for input node, we\ntrain HM2from two aspects: 1) corporate relative valuation\nloss, and 2) heterogeneous graph representation structural\nloss. This constructs a multi-task learning approach, which\ncan learn more discriminative representation. In result, the\noverall loss function is:\n\u2113=\u2113m+\u03bb\u2113b,\n\u2113m=\u2212\u2211\ni\u2208V1\u2211\nj1{yp\ni=j}logexp(\u03b8\u22a4\njf2(vi))\u2211\nkexp(\u03b8\u22a4\nkf2(vi)),\n\u2113b=\u2211\n<i,j,k>\u2208Tmax{0,\u00b5 +d(f 2(vi),f2(vj))\u2212d(f 2(vi),f2(vk))},\ns.t. yb\ni=yb\nj\u0338=yb\nk, yp\ni=yp\nj=yp\nk.\n(7)\nwhere\u2113mdenotes the corporate valuation loss, \u03b8is the\nfully connected layer to the prediction layer, \u00b5represents\nthe margin de\ufb01ned manually, and d(\u00b7) is the distance mea-\nsurement function, which measures the distance between\ntwo node embeddings (we utilize the euclidean distance\nhere for simplicity). Note that \u2113bre\ufb02ects the embedding ef-\nfect between competitors considered by traditional domain\nexperts, thus further regularizes the embedding structure\nin the same business category. Inspired by [26], we can\ncorporate the random walk in graph to generate triplets\n< i,j,k >\u2208T . In detail, we \ufb01rst generate a set of random\nwalks in the", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "e40e0d0c-2842-4989-9e31-f6cbe20c35ab": {"__data__": {"id_": "e40e0d0c-2842-4989-9e31-f6cbe20c35ab", "embedding": null, "metadata": {"page_label": "7", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of incorporating the random walk in graph to generate triplets in the Heterogeneous MultiModal Graph Neural Network for Corporate Relative Valuation?\n2. How does the proposed distance measurement function in the Heterogeneous MultiModal Graph Neural Network for Corporate Relative Valuation affect the embedding structure of competitors in the same business category?\n3. What is the procedure for training the HM2 model in the Heterogeneous MultiModal Graph Neural Network for Corporate Relative Valuation, and what are the key steps involved in the process?", "prev_section_summary": "The section discusses the use of a heterogeneous multi-modal graph neural network for corporate relative valuation. The model training process involves two aspects: corporate relative valuation loss and heterogeneous graph representation structural loss. The random walk in graph is used to generate triplets for the model training process. The overall loss function is a multi-task learning approach that can learn more discriminative representation. The section also mentions the distance measurement function used to measure the distance between node embeddings.", "section_summary": "The section discusses the use of a Heterogeneous MultiModal Graph Neural Network (HM2) for corporate relative valuation. The purpose of incorporating the random walk in graph to generate triplets in the HM2 model is to further regularize the embedding structure in the same business category. The proposed distance measurement function in the HM2 model affects the embedding structure of competitors in the same business category by reflecting the embedding effect between competitors considered by traditional domain experts. The procedure for training the HM2 model involves generating random walks in the HMMG, collecting node j and k with the same valuation level yp and different business category yb, calculating the objective according to Equation 7, and updating the model parameters via the Adam optimizer. The model can be used for inductive corporate relative valuation.", "excerpt_keywords": "1. HMMG\n2. node embeddings\n3. Euclidean distance\n4. random walk\n5. graph\n6. triplets\n7. Adam optimizer\n8. mini-batch\n9. classification loss\n10. triplet margin loss"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "b26ce5b7-5cd1-4c4f-ab00-7f0ec4074897", "node_type": "4", "metadata": {"page_label": "7", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "4a573f135fd6e0a296538270df39d45736b84875c1602c0db0139321f7e539ec"}, "2": {"node_id": "d8e60937-b046-4729-83f5-35b2bb6e3498", "node_type": "1", "metadata": {"page_label": "7", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "184e7dc8fbcce3e3a209af53417383f490d23a0694120a070295fb7af96aae2f"}, "3": {"node_id": "4bb4fea5-d8cc-4249-99cb-d5a06b1528ac", "node_type": "1", "metadata": {"page_label": "7", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "2ee34e2d61fe9cde19a09aee48db9ac93e9a8c5c0664db67c66354726a06d9b1"}}, "hash": "e4e94118d0faf899b4763ba56f9bc3dc8acb0139261fa52194a7885f1ac2151f", "text": "distance mea-\nsurement function, which measures the distance between\ntwo node embeddings (we utilize the euclidean distance\nhere for simplicity). Note that \u2113bre\ufb02ects the embedding ef-\nfect between competitors considered by traditional domain\nexperts, thus further regularizes the embedding structure\nin the same business category. Inspired by [26], we can\ncorporate the random walk in graph to generate triplets\n< i,j,k >\u2208T . In detail, we \ufb01rst generate a set of random\nwalks in the HMMG. Then we collect node jwith same\nthe same business category yb\niand valuation level yp\nifor\nnodeiin the walk sequence. Besides, we sample node\nkwith the same valuation level yp\nibut different business\ncategoryyb\nifor nodei. For optimizing HM2, we \ufb01rst sample\na mini-batch of triplets at each iteration, and calculate the\nobjective according to Equation 7. The model parameters\nare updated via the Adam optimizer [27]. And we utilize\nextra 10% randomly sampled data as early stop for better\ngeneralization. With the learned model, we can conduct\ninductive corporate relative valuation.\nThe procedure of training HM2model can be summa-\nrized in algorithm 1. Line 4 and line 5 correspond to our\nneighbour sampling module. Line 6 and Line 7 calculate\nthe proposed modal attribute encoding and the multi-modal\naggregating results respectively. Line 8 and line 9 calculate\nthe classi\ufb01cation loss and triplet margin loss respectively. In\neach epoch, HM2samples mini-batches of company nodes\nv1and update model parameters using gradient descent.Algorithm 1 The pseudo code of HM2\nInput:\n\u2022Dataset: HMMG (V,E,C V,CE), attribute x, ground\ntruth y;\n\u2022Parameter:\u03bb;\n\u2022maxIter:T, learning rate:", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "4bb4fea5-d8cc-4249-99cb-d5a06b1528ac": {"__data__": {"id_": "4bb4fea5-d8cc-4249-99cb-d5a06b1528ac", "embedding": null, "metadata": {"page_label": "7", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the proposed algorithm for corporate relative valuation using heterogeneous multi-modal graph neural networks, and how does it differ from state-of-the-art comparison methods?\n2. How do the different components of the proposed algorithm, such as modal attribute encoding and triplet margin loss, affect the estimation of corporate relative valuation?\n3. What are the hyperparameters used in the proposed algorithm, and how do they affect the performance of the model in estimating corporate relative valuation?", "prev_section_summary": "The section discusses the use of a Heterogeneous MultiModal Graph Neural Network (HM2) for corporate relative valuation. The purpose of incorporating the random walk in graph to generate triplets in the HM2 model is to further regularize the embedding structure in the same business category. The proposed distance measurement function in the HM2 model affects the embedding structure of competitors in the same business category by reflecting the embedding effect between competitors considered by traditional domain experts. The procedure for training the HM2 model involves generating random walks in the HMMG, collecting node j and k with the same valuation level yp and different business category yb, calculating the objective according to Equation 7, and updating the model parameters via the Adam optimizer. The model can be used for inductive corporate relative valuation.", "section_summary": "The section describes a proposed algorithm for corporate relative valuation using heterogeneous multi-modal graph neural networks. The algorithm involves modal attribute encoding, triplet margin loss, and gradient descent for model updates. The section also presents related experiments to validate the effectiveness of the proposed method. The experiments compare the performance of HM2 with state-of-the-art comparison methods and evaluate the impact of different components and hyperparameters on the estimation of corporate relative valuation.", "excerpt_keywords": "1. HM2, 2. state-of-the-art, 3. real-world dataset, 4. prediction performance, 5. components, 6. modal attribute encoding, 7. hyper-parameters, 8. estimation, 9. fairness, 10. positivity."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "b26ce5b7-5cd1-4c4f-ab00-7f0ec4074897", "node_type": "4", "metadata": {"page_label": "7", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "4a573f135fd6e0a296538270df39d45736b84875c1602c0db0139321f7e539ec"}, "2": {"node_id": "e40e0d0c-2842-4989-9e31-f6cbe20c35ab", "node_type": "1", "metadata": {"page_label": "7", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "e4e94118d0faf899b4763ba56f9bc3dc8acb0139261fa52194a7885f1ac2151f"}, "3": {"node_id": "60bde3cc-3c74-444d-808f-5d7a1bbcea2d", "node_type": "1", "metadata": {"page_label": "7", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "28621a41b00d85d0db1d6e97e70171ef0639bed459c4a538ade311acbd4ed3ad"}}, "hash": "2ee34e2d61fe9cde19a09aee48db9ac93e9a8c5c0664db67c66354726a06d9b1", "text": "and Line 7 calculate\nthe proposed modal attribute encoding and the multi-modal\naggregating results respectively. Line 8 and line 9 calculate\nthe classi\ufb01cation loss and triplet margin loss respectively. In\neach epoch, HM2samples mini-batches of company nodes\nv1and update model parameters using gradient descent.Algorithm 1 The pseudo code of HM2\nInput:\n\u2022Dataset: HMMG (V,E,C V,CE), attribute x, ground\ntruth y;\n\u2022Parameter:\u03bb;\n\u2022maxIter:T, learning rate: lr\nOutput:\n\u2022Classi\ufb01ers:F\n1:Initialize HM2model parameters \u0398;\n2:while stop condition is not triggered do\n3: formini-batch of company node v1do\n4: Gather neighbour nodes nfor each node in batch\nvia random walk;\n5: Select neighbour company and member nodes with\nhighest frequency, N1(v1)andN2(v1)respectively;\n6: Calculateft\n1(v1\ni)according to Equation 5;\n7: Calculatef2(v1\ni)according to Equation 6;\n8: Calculate\u2113m;\n9: Sample tripletsTand calculate \u2113b;\n10: Calculate loss \u2113=\u2113m+\u03bb\u2113baccording to Equation\n7;\n11: Update model parameters using gradient descent;\n12: end for\n13:end while\n4 E XPERIMENTS\nIn this section, we develop related experiments to validate\nthe effectiveness of our proposed method.\n\u2022Q1: How do HM2and state-of-the-art comparison\nmethods perform on real-world dataset? For example,\nthe prediction performances in De\ufb01nition 1.\n\u2022Q2:How do the components of HM2affect the estima-\ntion? For example, modal attribute encoding.\n\u2022Q3:How do various hyper-parameters in the approach\naffect", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "60bde3cc-3c74-444d-808f-5d7a1bbcea2d": {"__data__": {"id_": "60bde3cc-3c74-444d-808f-5d7a1bbcea2d", "embedding": null, "metadata": {"page_label": "7", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. How does the proposed heterogeneous multi-modal graph neural network (HM2) compare to traditional methods such as SVM, MLP, KNN, and state-of-the-art graph methods such as HetGNN, m2vec, ASNE, Graph-SAGE, GAT, HAN, and GATNE in terms of prediction performance on real-world datasets?\n2. What are the components of the HM2 model and how do they affect the estimation of corporate relative valuation? For example, what is the role of modal attribute encoding in the model?\n3. How do various hyperparameters in the HM2 approach affect its performance? For instance, what is the optimal size of sampled neighbors and embedding size for achieving the best results?", "prev_section_summary": "The section describes a proposed algorithm for corporate relative valuation using heterogeneous multi-modal graph neural networks. The algorithm involves modal attribute encoding, triplet margin loss, and gradient descent for model updates. The section also presents related experiments to validate the effectiveness of the proposed method. The experiments compare the performance of HM2 with state-of-the-art comparison methods and evaluate the impact of different components and hyperparameters on the estimation of corporate relative valuation.", "section_summary": "The section discusses the proposed heterogeneous multi-modal graph neural network (HM2) for corporate relative valuation and compares its performance to traditional methods such as SVM, MLP, KNN, and state-of-the-art graph methods such as HetGNN, m2vec, ASNE, Graph-SAGE, GAT, HAN, and GATNE. The section also discusses the components of the HM2 model and how they affect the estimation of corporate relative valuation, as well as how various hyperparameters in the approach affect its performance. The section presents related experiments to validate the effectiveness of the proposed method, including comparisons with other methods on real-world datasets.", "excerpt_keywords": "1. HM2, 2. state-of-the-art comparison, 3. real-world dataset, 4. prediction performances, 5. Definition 1, 6. components of HM2, 7. estimation, 8. modal attribute encoding, 9. hyper-parameters, 10. SVM, MLP, KNN, HetGNN, m2vec, ASNE, Graph-SAGE, GAT, HAN, GATNE."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "b26ce5b7-5cd1-4c4f-ab00-7f0ec4074897", "node_type": "4", "metadata": {"page_label": "7", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "4a573f135fd6e0a296538270df39d45736b84875c1602c0db0139321f7e539ec"}, "2": {"node_id": "4bb4fea5-d8cc-4249-99cb-d5a06b1528ac", "node_type": "1", "metadata": {"page_label": "7", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "2ee34e2d61fe9cde19a09aee48db9ac93e9a8c5c0664db67c66354726a06d9b1"}, "3": {"node_id": "f961527e-ce85-43f9-a7fc-dbf858464d20", "node_type": "1", "metadata": {"page_label": "7", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "7710aaccdb83f10739596307bbd8d1f5520721f78384f02a4f27711d2231d803"}}, "hash": "28621a41b00d85d0db1d6e97e70171ef0639bed459c4a538ade311acbd4ed3ad", "text": "end for\n13:end while\n4 E XPERIMENTS\nIn this section, we develop related experiments to validate\nthe effectiveness of our proposed method.\n\u2022Q1: How do HM2and state-of-the-art comparison\nmethods perform on real-world dataset? For example,\nthe prediction performances in De\ufb01nition 1.\n\u2022Q2:How do the components of HM2affect the estima-\ntion? For example, modal attribute encoding.\n\u2022Q3:How do various hyper-parameters in the approach\naffect performance? For example, the size of sampled\nneighbor and the embedding size.\nWe compare our HM2model with three traditional\nmethods: SVM, MLP , KNN, and seven state-of-the-art graph\nmethods: HetGNN [19], m2vec [28], ASNE [29], Graph-\nSAGE (SAGE for simplicity) [18], GAT [25], HAN [30], and\nGATNE [31]. All the methods are given the best perfor-\nmance as [19]. The details are:\n\u2022SVM: A linear method that considers single modal\nfeatures as input, in detail, we develop the attributes\nof corporate node as the input;\n\u2022MLP: A fully connected network that considers single\nmodal features as input, in detail, we develop the\nattributes of corporate node as the input;\n\u2022HetGNN: A heterogeneous graph neural network\nmodel that constructs two modules to aggregate feature\ninformation of heterogeneous nodes respectively, in\nwhich the \ufb01rst module learns embeddings of heteroge-\nneous contents with the LSTM module, and the second\nmodule aggregates embeddings of different neighbor-\ning types for obtaining the \ufb01nal node embedding [13];\n\u2022m2vec: A heterogeneous graph model that leverages\nmeta-path based random walks in heterogeneous net-\nworks to generate heterogeneous neighborhoods, then\nAuthorized licensed use limited to: Univ", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "f961527e-ce85-43f9-a7fc-dbf858464d20": {"__data__": {"id_": "f961527e-ce85-43f9-a7fc-dbf858464d20", "embedding": null, "metadata": {"page_label": "7", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the heterogeneous graph neural network model described in the paper \"Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf\"?\n2. How does the heterogeneous graph neural network model construct the two modules to aggregate feature information of heterogeneous nodes?\n3. What is the difference between the heterogeneous graph neural network model and the m2vec model in terms of their approach to generating heterogeneous neighborhoods?", "prev_section_summary": "The section discusses the proposed heterogeneous multi-modal graph neural network (HM2) for corporate relative valuation and compares its performance to traditional methods such as SVM, MLP, KNN, and state-of-the-art graph methods such as HetGNN, m2vec, ASNE, Graph-SAGE, GAT, HAN, and GATNE. The section also discusses the components of the HM2 model and how they affect the estimation of corporate relative valuation, as well as how various hyperparameters in the approach affect its performance. The section presents related experiments to validate the effectiveness of the proposed method, including comparisons with other methods on real-world datasets.", "section_summary": "The section discusses a heterogeneous graph neural network model for corporate relative valuation. The model constructs two modules to aggregate feature information of heterogeneous nodes, with the first module learning embeddings of heterogeneous contents using an LSTM module, and the second module aggregating embeddings of different neighboring types. The section also compares the heterogeneous graph neural network model to the m2vec model in terms of their approach to generating heterogeneous neighborhoods.", "excerpt_keywords": "heterogeneous graph, neural network, LSTM, embedding, heterogeneous nodes, graph neural network, graph model, meta-path, random walks, heterogeneous neighborhoods."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "b26ce5b7-5cd1-4c4f-ab00-7f0ec4074897", "node_type": "4", "metadata": {"page_label": "7", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "4a573f135fd6e0a296538270df39d45736b84875c1602c0db0139321f7e539ec"}, "2": {"node_id": "60bde3cc-3c74-444d-808f-5d7a1bbcea2d", "node_type": "1", "metadata": {"page_label": "7", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "28621a41b00d85d0db1d6e97e70171ef0639bed459c4a538ade311acbd4ed3ad"}}, "hash": "7710aaccdb83f10739596307bbd8d1f5520721f78384f02a4f27711d2231d803", "text": "A heterogeneous graph neural network\nmodel that constructs two modules to aggregate feature\ninformation of heterogeneous nodes respectively, in\nwhich the \ufb01rst module learns embeddings of heteroge-\nneous contents with the LSTM module, and the second\nmodule aggregates embeddings of different neighbor-\ning types for obtaining the \ufb01nal node embedding [13];\n\u2022m2vec: A heterogeneous graph model that leverages\nmeta-path based random walks in heterogeneous net-\nworks to generate heterogeneous neighborhoods, then\nAuthorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 20,2021 at 08:36:01 UTC from IEEE Xplore.  Restrictions apply.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "035a0bb6-ce4b-4b04-bf6e-bda7e200b914": {"__data__": {"id_": "035a0bb6-ce4b-4b04-bf6e-bda7e200b914", "embedding": null, "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the accuracy of the Heterogeneous Multi-Modal Graph Neural Network (HetGNN) in predicting corporate relative valuation for different training data ratios?\n2. How does the performance of the HetGNN compare to other machine learning algorithms such as Support Vector Machines (SVM), Multi-Layer Perceptron (MLP), and K-Nearest Neighbors (KNN) in predicting corporate relative valuation?\n3. What is the precision of the HetGNN in predicting corporate relative valuation for different training data ratios?", "prev_section_summary": "The section discusses a heterogeneous graph neural network model for corporate relative valuation. The model constructs two modules to aggregate feature information of heterogeneous nodes, with the first module learning embeddings of heterogeneous contents using an LSTM module, and the second module aggregating embeddings of different neighboring types. The section also compares the heterogeneous graph neural network model to the m2vec model in terms of their approach to generating heterogeneous neighborhoods.", "section_summary": "The section discusses the accuracy and performance of a Heterogeneous Multi-Modal Graph Neural Network (HetGNN) in predicting corporate relative valuation. The study compares the performance of HetGNN to other machine learning algorithms such as Support Vector Machines (SVM), Multi-Layer Perceptron (MLP), and K-Nearest Neighbors (KNN) in predicting corporate relative valuation. The study also examines the precision of HetGNN in predicting corporate relative valuation for different training data ratios. The results are presented in Table 3, which shows the accuracy and precision of each algorithm for different training data ratios.", "excerpt_keywords": "1. Corporate relative valuation prediction\n2. SVM\n3. MLP\n4. KNN\n5. HetGNN\n6. GAT\n7. SAGE\n8. ASNE\n9. m2vec\n10. HAN"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "8500b2d6-015e-436e-a558-aa4efc073e92", "node_type": "4", "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "9a57af184d633ae88d28a9adf02309b9bec7408e52023bc5e2801fb60d74ef6e"}, "3": {"node_id": "eda5aba7-a3f1-4bf9-9dd3-eaa15059a4aa", "node_type": "1", "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a54f2eab654e2e91065e4ad575d49c34ba4f8cedb8ff7442260587f81faf2044"}}, "hash": "e61a0b877ff01c0a3ded5db8c19fb532c42d33db9f1b3fdb98bba3e3df80145d", "text": "1041-4347 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TKDE.2021.3080293, IEEE\nTransactions on Knowledge and Data Engineering\n8\nTABLE 3\nCorporate relative valuation prediction results (y ), percentage denotes training data ratio. The best results are highlighted in bold.\nMetric SVM MLP KNN HetGNN GAT SAGE ASNE m2vec HAN GATNE FAME HM2\nAccuracy10% .273\u00b1.002 .247\u00b1.014 .326\u00b1.012 .307\u00b1.004 .313\u00b1.011 .304\u00b1.012 .289\u00b1.011 .299\u00b1.008 .310\u00b1.011 .304\u00b1.004 .322\u00b1.014 .346\u00b1.008\n30% .305\u00b1.003 .305\u00b1.047 .333\u00b1.005 .353\u00b1.009 .346\u00b1.008 .364\u00b1.011 .359\u00b1.002 .342\u00b1.010 .347\u00b1.011 .330\u00b1.004 .349\u00b1.002 .388\u00b1.004\n50% .336\u00b1.002 .340\u00b1.046 .347\u00b1.009 .377\u00b1.005 .378\u00b1.003 .395\u00b1.007 .387\u00b1.007 .360\u00b1.002 .393\u00b1.009 .346\u00b1.002 .380\u00b1.005 .410\u00b1.008\n70% .367\u00b1.002 .374\u00b1.029 .349\u00b1.010 .393\u00b1.008 .407\u00b1.009 .413\u00b1.010 .399\u00b1.003 .388\u00b1.016 .403\u00b1.009 .357\u00b1.007 .390\u00b1.010 .446\u00b1.007\nPrecision10% .286\u00b1.003 .172\u00b1.097 .331\u00b1.010 .301\u00b1.009", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "eda5aba7-a3f1-4bf9-9dd3-eaa15059a4aa": {"__data__": {"id_": "eda5aba7-a3f1-4bf9-9dd3-eaa15059a4aa", "embedding": null, "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the performance of the heterogeneous multi-modal graph neural network in predicting corporate relative valuation?\n2. How does the precision and recall of the model vary across different categories of companies?\n3. What is the impact of the heterogeneity of the input data on the performance of the model?", "prev_section_summary": "The section discusses the accuracy and performance of a Heterogeneous Multi-Modal Graph Neural Network (HetGNN) in predicting corporate relative valuation. The study compares the performance of HetGNN to other machine learning algorithms such as Support Vector Machines (SVM), Multi-Layer Perceptron (MLP), and K-Nearest Neighbors (KNN) in predicting corporate relative valuation. The study also examines the precision of HetGNN in predicting corporate relative valuation for different training data ratios. The results are presented in Table 3, which shows the accuracy and precision of each algorithm for different training data ratios.", "section_summary": "The section discusses the performance of a heterogeneous multi-modal graph neural network in predicting corporate relative valuation. The model's precision and recall vary across different categories of companies, and the heterogeneity of the input data impacts the performance of the model. The section provides tables with the precision, recall, and F1-score for different categories of companies and the overall performance of the model.", "excerpt_keywords": "1. Assist, care, respect, truth, utility, security, harmful, unethical, prejudiced, negative, fairness, positivity, precision, recall, percentage, confidence interval, standard deviation, mean, median, mode, range, variance."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "8500b2d6-015e-436e-a558-aa4efc073e92", "node_type": "4", "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "9a57af184d633ae88d28a9adf02309b9bec7408e52023bc5e2801fb60d74ef6e"}, "2": {"node_id": "035a0bb6-ce4b-4b04-bf6e-bda7e200b914", "node_type": "1", "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "e61a0b877ff01c0a3ded5db8c19fb532c42d33db9f1b3fdb98bba3e3df80145d"}, "3": {"node_id": "ab216420-b362-48ea-8290-5b42c8350509", "node_type": "1", "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "49262ce4c7c49a1a1e93635325b09b5a6275c4c2709297bc241f4e88d567a2b6"}}, "hash": "a54f2eab654e2e91065e4ad575d49c34ba4f8cedb8ff7442260587f81faf2044", "text": ".395\u00b1.007 .387\u00b1.007 .360\u00b1.002 .393\u00b1.009 .346\u00b1.002 .380\u00b1.005 .410\u00b1.008\n70% .367\u00b1.002 .374\u00b1.029 .349\u00b1.010 .393\u00b1.008 .407\u00b1.009 .413\u00b1.010 .399\u00b1.003 .388\u00b1.016 .403\u00b1.009 .357\u00b1.007 .390\u00b1.010 .446\u00b1.007\nPrecision10% .286\u00b1.003 .172\u00b1.097 .331\u00b1.010 .301\u00b1.009 .319\u00b1.313 .305\u00b1.013 .290\u00b1.014 .294\u00b1.010 .309\u00b1.011 .304\u00b1.003 .323\u00b1.011 .351\u00b1.006\n30% .341\u00b1.006 .266\u00b1.095 .337\u00b1.004 .355\u00b1.011 .339\u00b1.010 .354\u00b1.011 .361\u00b1.004 .335\u00b1.008 .346\u00b1.007 .329\u00b1.005 .351\u00b1.004 .395\u00b1.004\n50% .334\u00b1.005 .292\u00b1.102 .351\u00b1.009 .386\u00b1.007 .375\u00b1.003 .386\u00b1.007 .386\u00b1.013 .362\u00b1.005 .386\u00b1.014 .344\u00b1.003 .377\u00b1.007 .405\u00b1.012\n70% .358\u00b1.004 .384\u00b1.021 .355\u00b1.013 .385\u00b1.012 .404\u00b1.009 .408\u00b1.010 .400\u00b1.002 .385\u00b1.016 .398\u00b1.013 .354\u00b1.006 .383\u00b1.011 .428\u00b1.004\nRecall10% .273\u00b1.007 .247\u00b1.014 .326\u00b1.012 .307\u00b1.004 .313\u00b1.011 .304\u00b1.015 .289\u00b1.011 .299\u00b1.008 .310\u00b1.011 .304\u00b1.004 .322\u00b1.014 .346\u00b1.008\n30% .305\u00b1.004 .305\u00b1.047 .333\u00b1.005 .353\u00b1.009 .346\u00b1.010 .364\u00b1.011", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "ab216420-b362-48ea-8290-5b42c8350509": {"__data__": {"id_": "ab216420-b362-48ea-8290-5b42c8350509", "embedding": null, "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the performance of the heterogeneous multi-modal graph neural network in predicting corporate relative valuation using 10%, 30%, 50%, and 70% recall?\n2. How does the heterogeneous multi-modal graph neural network compare to other models in terms of F1-measure for predicting corporate relative valuation?\n3. What are the key features or characteristics of the heterogeneous multi-modal graph neural network that contribute to its high performance in predicting corporate relative valuation?", "prev_section_summary": "The section discusses the performance of a heterogeneous multi-modal graph neural network in predicting corporate relative valuation. The model's precision and recall vary across different categories of companies, and the heterogeneity of the input data impacts the performance of the model. The section provides tables with the precision, recall, and F1-score for different categories of companies and the overall performance of the model.", "section_summary": "The section discusses the performance of a heterogeneous multi-modal graph neural network in predicting corporate relative valuation using 10%, 30%, 50%, and 70% recall. The network is compared to other models in terms of F1-measure for predicting corporate relative valuation. The key features or characteristics of the heterogeneous multi-modal graph neural network that contribute to its high performance in predicting corporate relative valuation are not explicitly stated in the provided excerpt.", "excerpt_keywords": "care, respect, truth, utility, security, ethical, prejudiced, negative, fairness, positivity, recall, F1-measure, 10%, 30%, 50%, 70%, 100%, 10%, 30%, 50%, 70%, 100%, 10%, 30%, 50%, 70%, 100%, 10%, 30%, 50%, 70%, 100%, 10%, 30%, 50%, 70%, 100%, 10%, 30%, 50%, 70%, 100%, 10%, 30%, 50%, 70%, 100%, 10%, 30%, 50%, 70%, 100%, 10%, 30%, 50%, 70%, 100%, 10%, 30%, 50%, 70%, 100%, 10%, 30%, 50%, 70%,"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "8500b2d6-015e-436e-a558-aa4efc073e92", "node_type": "4", "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "9a57af184d633ae88d28a9adf02309b9bec7408e52023bc5e2801fb60d74ef6e"}, "2": {"node_id": "eda5aba7-a3f1-4bf9-9dd3-eaa15059a4aa", "node_type": "1", "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a54f2eab654e2e91065e4ad575d49c34ba4f8cedb8ff7442260587f81faf2044"}, "3": {"node_id": "e9a91b17-9249-41b8-8ec9-c8c477194fcb", "node_type": "1", "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "f00e68ea28e28e2cdb7303198ba59dee01846a87a071ae643efd48278c5fe116"}}, "hash": "49262ce4c7c49a1a1e93635325b09b5a6275c4c2709297bc241f4e88d567a2b6", "text": ".400\u00b1.002 .385\u00b1.016 .398\u00b1.013 .354\u00b1.006 .383\u00b1.011 .428\u00b1.004\nRecall10% .273\u00b1.007 .247\u00b1.014 .326\u00b1.012 .307\u00b1.004 .313\u00b1.011 .304\u00b1.015 .289\u00b1.011 .299\u00b1.008 .310\u00b1.011 .304\u00b1.004 .322\u00b1.014 .346\u00b1.008\n30% .305\u00b1.004 .305\u00b1.047 .333\u00b1.005 .353\u00b1.009 .346\u00b1.010 .364\u00b1.011 .359\u00b1.002 .338\u00b1.010 .347\u00b1.011 .330\u00b1.004 .349\u00b1.002 .388\u00b1.004\n50% .336\u00b1.006 .340\u00b1.046 .346\u00b1.009 .377\u00b1.005 .378\u00b1.003 .395\u00b1.007 .384\u00b1.007 .358\u00b1.002 .393\u00b1.009 .346\u00b1.002 .380\u00b1.005 .410\u00b1.012\n70% .367\u00b1.005 .374\u00b1.029 .349\u00b1.010 .393\u00b1.008 .407\u00b1.009 .413\u00b1.010 .399\u00b1.003 .388\u00b1.016 .403\u00b1.009 .357\u00b1.007 .390\u00b1.010 .446\u00b1.005\nF1-measure10% .175\u00b1.003 .122\u00b1.038 .326\u00b1.011 .301\u00b1.010 .313\u00b1.011 .302\u00b1.012 .288\u00b1.013 .292\u00b1.016 .301\u00b1.010 .297\u00b1.009 .322\u00b1.013 .340\u00b1.009\n30% .269\u00b1.004 .220\u00b1.085 .334\u00b1.004 .345\u00b1.016 .339\u00b1.008 .355\u00b1.012 .359\u00b1.003 .335\u00b1.012 .334\u00b1.018 .328\u00b1.005 .349\u00b1.004 .376\u00b1.005\n50% .327\u00b1.003 .275\u00b1.085 .347\u00b1.009 .377\u00b1.006 .376\u00b1.003 .389\u00b1.007", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "e9a91b17-9249-41b8-8ec9-c8c477194fcb": {"__data__": {"id_": "e9a91b17-9249-41b8-8ec9-c8c477194fcb", "embedding": null, "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the performance of different machine learning algorithms in predicting corporate relative valuation using heterogeneous multi-modal graph neural networks?\n2. How does the incorporation of heterogeneous multi-modal data improve the accuracy of corporate relative valuation predictions compared to traditional methods?\n3. What are the key factors that influence the effectiveness of heterogeneous multi-modal graph neural networks in predicting corporate relative valuation?", "prev_section_summary": "The section discusses the performance of a heterogeneous multi-modal graph neural network in predicting corporate relative valuation using 10%, 30%, 50%, and 70% recall. The network is compared to other models in terms of F1-measure for predicting corporate relative valuation. The key features or characteristics of the heterogeneous multi-modal graph neural network that contribute to its high performance in predicting corporate relative valuation are not explicitly stated in the provided excerpt.", "section_summary": "The section discusses the use of heterogeneous multi-modal graph neural networks (HetGNN) for predicting corporate relative valuation. The performance of different machine learning algorithms, including support vector machines (SVM), multi-layer perceptron (MLP), k-nearest neighbors (KNN), and graph neural networks (GNN), is compared. The incorporation of heterogeneous multi-modal data improves the accuracy of corporate relative valuation predictions compared to traditional methods. The key factors that influence the effectiveness of HetGNN in predicting corporate relative valuation include the type of data used, the architecture of the network, and the training data ratio. The results of the study are presented in Table 4, which shows the mean squared error (MSE) for different metrics and training data ratios.", "excerpt_keywords": "1. Corporate relative valuation, 2. SVM, 3. MLP, 4. KNN, 5. HetGNN, 6. GAT, 7. SAGE, 8. ASNE, 9. m2vec, 10. HAN."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "8500b2d6-015e-436e-a558-aa4efc073e92", "node_type": "4", "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "9a57af184d633ae88d28a9adf02309b9bec7408e52023bc5e2801fb60d74ef6e"}, "2": {"node_id": "ab216420-b362-48ea-8290-5b42c8350509", "node_type": "1", "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "49262ce4c7c49a1a1e93635325b09b5a6275c4c2709297bc241f4e88d567a2b6"}, "3": {"node_id": "c5b7cc46-849c-465a-8715-8d69c0f9334d", "node_type": "1", "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a644b7925325bc94a6b0c91b65bb479118c8667e1838b2ba9d99eb7a67432d18"}}, "hash": "f00e68ea28e28e2cdb7303198ba59dee01846a87a071ae643efd48278c5fe116", "text": ".288\u00b1.013 .292\u00b1.016 .301\u00b1.010 .297\u00b1.009 .322\u00b1.013 .340\u00b1.009\n30% .269\u00b1.004 .220\u00b1.085 .334\u00b1.004 .345\u00b1.016 .339\u00b1.008 .355\u00b1.012 .359\u00b1.003 .335\u00b1.012 .334\u00b1.018 .328\u00b1.005 .349\u00b1.004 .376\u00b1.005\n50% .327\u00b1.003 .275\u00b1.085 .347\u00b1.009 .377\u00b1.006 .376\u00b1.003 .389\u00b1.007 .381\u00b1.009 .359\u00b1.004 .385\u00b1.010 .343\u00b1.003 .375\u00b1.009 .400\u00b1.008\n70% .351\u00b1.004 .333\u00b1.053 .350\u00b1.011 .381\u00b1.015 .405\u00b1.009 .410\u00b1.010 .398\u00b1.003 .385\u00b1.017 .397\u00b1.013 .351\u00b1.007 .384\u00b1.011 .424\u00b1.007\nTABLE 4\nCorporate relative valuation results ( \u02c6y), percentage denotes training data ratio. The best results are highlighted in bold.\nMetric SVM MLP KNN HetGNN GAT SAGE ASNE m2vec HAN GATNE FAME HM2\nMSE10% 4.277\u00b1.057 4.268\u00b1.212 4.188\u00b1.077 4.122\u00b1.033 5.247\u00b1.304 4.998\u00b1.176 3.822\u00b1.080 4.599\u00b1.081 4.496\u00b1.198 4.927\u00b1.060 4.796\u00b1.093 3.919\u00b1.077\n30% 4.029\u00b1.104 4.170\u00b1.250 4.165\u00b1.093 4.002\u00b1.046 4.225\u00b1.080 4.517\u00b1.096 3.823\u00b1.056 4.458\u00b1.024 4.207\u00b1.116 4.729\u00b1.029 4.607\u00b1.059 3.481\u00b1.051\n50%", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "c5b7cc46-849c-465a-8715-8d69c0f9334d": {"__data__": {"id_": "c5b7cc46-849c-465a-8715-8d69c0f9334d", "embedding": null, "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation?\n2. How does the model preserve both the structures and semantics of the given heterogeneous network?\n3. What is the difference between the ASNE, GraphSAGE, and skip-gram models used in the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation?", "prev_section_summary": "The section discusses the use of heterogeneous multi-modal graph neural networks (HetGNN) for predicting corporate relative valuation. The performance of different machine learning algorithms, including support vector machines (SVM), multi-layer perceptron (MLP), k-nearest neighbors (KNN), and graph neural networks (GNN), is compared. The incorporation of heterogeneous multi-modal data improves the accuracy of corporate relative valuation predictions compared to traditional methods. The key factors that influence the effectiveness of HetGNN in predicting corporate relative valuation include the type of data used, the architecture of the network, and the training data ratio. The results of the study are presented in Table 4, which shows the mean squared error (MSE) for different metrics and training data ratios.", "section_summary": "The section discusses the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation, which is a model used to preserve both the structures and semantics of a given heterogeneous network. The model uses three different techniques: ASNE, GraphSAGE, and skip-gram, to learn representations for nodes while preserving structural and attribute proximity. The model extends the skip-gram model to facilitate the modeling of connected nodes. The section also provides some performance metrics for the model.", "excerpt_keywords": "1. Graph neural networks, 2. Node embeddings, 3. Heterogeneous networks, 4. Structural proximity, 5. Attribute proximity, 6. Skip-gram model, 7. Attributed graph embedding, 8. Inductive graph neural network, 9. Sampling, 10. Aggregating features"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "8500b2d6-015e-436e-a558-aa4efc073e92", "node_type": "4", "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "9a57af184d633ae88d28a9adf02309b9bec7408e52023bc5e2801fb60d74ef6e"}, "2": {"node_id": "e9a91b17-9249-41b8-8ec9-c8c477194fcb", "node_type": "1", "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "f00e68ea28e28e2cdb7303198ba59dee01846a87a071ae643efd48278c5fe116"}, "3": {"node_id": "bd4d0ef9-9a64-4eca-9203-80315058db15", "node_type": "1", "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "3c3419932d293a628a6a2f48c8b8fbb6eab31381a242fdf65d04bb4bfddde364"}}, "hash": "a644b7925325bc94a6b0c91b65bb479118c8667e1838b2ba9d99eb7a67432d18", "text": "3.822\u00b1.080 4.599\u00b1.081 4.496\u00b1.198 4.927\u00b1.060 4.796\u00b1.093 3.919\u00b1.077\n30% 4.029\u00b1.104 4.170\u00b1.250 4.165\u00b1.093 4.002\u00b1.046 4.225\u00b1.080 4.517\u00b1.096 3.823\u00b1.056 4.458\u00b1.024 4.207\u00b1.116 4.729\u00b1.029 4.607\u00b1.059 3.481\u00b1.051\n50% 3.818\u00b1.098 3.856\u00b1.114 4.143\u00b1.115 3.868\u00b1.072 3.872\u00b1.116 4.366\u00b1.207 3.638\u00b1.065 4.247\u00b1.051 3.970\u00b1.096 4.633\u00b1.022 4.561\u00b1.134 3.432\u00b1.071\n70% 3.523\u00b1.130 3.680\u00b1.219 4.127\u00b1.108 3.723\u00b1.039 3.453\u00b1.127 3.981\u00b1.043 3.487\u00b1.080 4.005\u00b1.045 3.858\u00b1.304 4.622\u00b1.030 4.475\u00b1.163 2.951\u00b1.084\nextends the skip-gram model to facilitate the modeling\nof connected nodes. The model can preserve both the\nstructures and semantics of the given heterogeneous\nnetwork by maximizing the likelihood [28];\n\u2022ASNE: An attributed graph embedding method, which\nlearns representations for nodes by preserving both\nthe structural proximity (capturing the global network\nstructure) and attribute proximity [29];\n\u2022GraphSAGE (SAGE for simplicity): An inductive\ngraph neural network model that leverages node fea-\nture to ef\ufb01ciently generate node embeddings for un-\nseen data. The model generates embeddings through\nsampling and aggregating features from a node\u2019s lo-\ncal neighborhood by different neural networks,", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "bd4d0ef9-9a64-4eca-9203-80315058db15": {"__data__": {"id_": "bd4d0ef9-9a64-4eca-9203-80315058db15", "embedding": null, "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using heterogeneous multi-modal graph neural networks for corporate relative valuation?\n2. How do different graph neural network models, such as GraphSAGE, GAT, and HAN, differ in their approach to generating node embeddings?\n3. What are the advantages and disadvantages of using unsupervised graph node embedding learning methods, such as HetGNN, ASNE, and m2vec, for corporate relative valuation?", "prev_section_summary": "The section discusses the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation, which is a model used to preserve both the structures and semantics of a given heterogeneous network. The model uses three different techniques: ASNE, GraphSAGE, and skip-gram, to learn representations for nodes while preserving structural and attribute proximity. The model extends the skip-gram model to facilitate the modeling of connected nodes. The section also provides some performance metrics for the model.", "section_summary": "The section discusses the use of heterogeneous multi-modal graph neural networks for corporate relative valuation. The section explains the different graph neural network models, such as GraphSAGE, GAT, and HAN, and their approach to generating node embeddings. The section also discusses the advantages and disadvantages of using unsupervised graph node embedding learning methods, such as HetGNN, ASNE, and m2vec, for corporate relative valuation. The section provides details on the implementation of these methods, including the source code provided by the authors and the modifications made to the dataset to conform to their input formats.", "excerpt_keywords": "graph neural networks, node embeddings, graph embedding methods, ASNE, GraphSAGE, GAT, HAN, GATNE, FAME, HetGNN, unsupervised learning, meta-paths, walk length, content features, general features, latent space, attribute semantics, multi-type relations, network structure, attribute proximity, structural proximity, self-attention, neighborhood information."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "8500b2d6-015e-436e-a558-aa4efc073e92", "node_type": "4", "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "9a57af184d633ae88d28a9adf02309b9bec7408e52023bc5e2801fb60d74ef6e"}, "2": {"node_id": "c5b7cc46-849c-465a-8715-8d69c0f9334d", "node_type": "1", "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a644b7925325bc94a6b0c91b65bb479118c8667e1838b2ba9d99eb7a67432d18"}, "3": {"node_id": "f8b751bb-40f9-4123-b6fc-5575dc5f2383", "node_type": "1", "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "1bbf85a6280a445b8b5e3bf5c1904529d0be184df24047edc1725801ef329848"}}, "hash": "3c3419932d293a628a6a2f48c8b8fbb6eab31381a242fdf65d04bb4bfddde364", "text": "by maximizing the likelihood [28];\n\u2022ASNE: An attributed graph embedding method, which\nlearns representations for nodes by preserving both\nthe structural proximity (capturing the global network\nstructure) and attribute proximity [29];\n\u2022GraphSAGE (SAGE for simplicity): An inductive\ngraph neural network model that leverages node fea-\nture to ef\ufb01ciently generate node embeddings for un-\nseen data. The model generates embeddings through\nsampling and aggregating features from a node\u2019s lo-\ncal neighborhood by different neural networks, i.e.,\nLSTM [18];\n\u2022GAT: A graph network model, which aggregates neigh-\nbors\u2019 information by masked self-attentions [25];\n\u2022HAN: A novel heterogeneous graph neural network\nbased on the hierarchical attention, including node-\nlevel and semantic-level attentions;\n\u2022GATNE: A heterogeneous network that splits the over-\nall node embedding into three parts: base, edge, and\nattribute embedding. The base embedding and attribute\nembedding are shared among edges of different types,\nwhile the edge embedding is computed by aggregation\nof neighborhood information with the self-attention\nmechanism;\n\u2022FAME: A heterogeneous network that maps the units\nfrom different modalities into the same latent space,\nwhich can preserve both attribute semantics and multi-\ntype relations in the learned embeddings.HetGNN, ASNE and m2vec are unsupervised graph node\nembedding learning methods. Thereby we use the source\ncode provided by the authors, and modify our dataset to\nconform to their input formats. For m2vec, we employ three\nmeta-paths, i.e., company-company, company-member-\ncompany, and company-member-member-company respec-\ntively. In addition, the walk length is set to 300. For ASNE,\nwe employ the same content features of different modalities\nas HM2and concatenate them as general features besides\nthe", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "f8b751bb-40f9-4123-b6fc-5575dc5f2383": {"__data__": {"id_": "f8b751bb-40f9-4123-b6fc-5575dc5f2383", "embedding": null, "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What are the three meta-paths used in the m2vec unsupervised graph node embedding learning method for corporate relative valuation?\n2. How are the input features and sampled neighbors set used in GraphSAGE and GAT for corporate relative valuation compared to HM2?\n3. What is the role of the validation set in tuning hyperparameters for the Heterogeneous Multi-Modal Graph Neural Network for corporate relative valuation?", "prev_section_summary": "The section discusses the use of heterogeneous multi-modal graph neural networks for corporate relative valuation. The section explains the different graph neural network models, such as GraphSAGE, GAT, and HAN, and their approach to generating node embeddings. The section also discusses the advantages and disadvantages of using unsupervised graph node embedding learning methods, such as HetGNN, ASNE, and m2vec, for corporate relative valuation. The section provides details on the implementation of these methods, including the source code provided by the authors and the modifications made to the dataset to conform to their input formats.", "section_summary": "The section discusses the use of Heterogeneous Multi-Modal Graph Neural Network (HM2) for corporate relative valuation. The authors use three meta-paths for m2vec unsupervised graph node embedding learning method, and the same input features and sampled neighbors set for GraphSAGE and GAT. The validation set is used to tune hyperparameters for the HM2 model. The section provides details on the implementation of the model, including the number of embedding aggregation heads, the dimension of edge and node embeddings, and the batch size. The authors also mention the use of cross-validation to tune the hyperparameters and the early stop criterion for comparing methods.", "excerpt_keywords": "graph node embedding, unsupervised learning, ASNE, GraphSAGE, GAT, m2vec, content features, latent features, early stop criterion, hyper-parameters, cross validation, validation set, training set, neural networks, classification, regression, node embedding, edge embedding, walk length, probability of returning to starting point, optimization, deep learning, graph learning, graph representation learning, graph neural networks, graph embedding learning, graph embedding optimization, graph embedding tuning, graph embedding selection, graph embedding evaluation, graph embedding comparison, graph embedding analysis, graph embedding optimization, graph embedding selection, graph embedding evaluation, graph embedding comparison, graph embedding analysis, graph embedding optimization, graph embedding selection, graph embedding evaluation, graph embedding comparison, graph embedding analysis, graph embedding optimization, graph embedding selection, graph embedding evaluation, graph embedding comparison, graph embedding analysis, graph embedding optimization, graph embedding selection, graph embedding evaluation, graph embedding comparison, graph embedding analysis, graph embedding optimization, graph embedding selection, graph embedding evaluation, graph embedding comparison, graph embedding analysis, graph embedding optimization, graph embedding selection, graph embedding evaluation, graph embedding comparison, graph embedding analysis, graph embedding optimization, graph embedding selection, graph embedding evaluation, graph embedding comparison, graph embedding"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "8500b2d6-015e-436e-a558-aa4efc073e92", "node_type": "4", "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "9a57af184d633ae88d28a9adf02309b9bec7408e52023bc5e2801fb60d74ef6e"}, "2": {"node_id": "bd4d0ef9-9a64-4eca-9203-80315058db15", "node_type": "1", "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "3c3419932d293a628a6a2f48c8b8fbb6eab31381a242fdf65d04bb4bfddde364"}, "3": {"node_id": "c7ae3fa8-e717-4a52-9b40-c15d32848a2b", "node_type": "1", "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "286c256e8b51d6ef7260935ec604419d15f0772ed059c55e83119c162de7233a"}}, "hash": "1bbf85a6280a445b8b5e3bf5c1904529d0be184df24047edc1725801ef329848", "text": "and m2vec are unsupervised graph node\nembedding learning methods. Thereby we use the source\ncode provided by the authors, and modify our dataset to\nconform to their input formats. For m2vec, we employ three\nmeta-paths, i.e., company-company, company-member-\ncompany, and company-member-member-company respec-\ntively. In addition, the walk length is set to 300. For ASNE,\nwe employ the same content features of different modalities\nas HM2and concatenate them as general features besides\nthe latent features. For GraphSAGE and GAT, we use the\nsame input features and sampled neighbors set for each\nnode as HM2. With the learned embeddings, we train an\nMLP model to obtain a classi\ufb01er or regressor as HM2. GAT\nand SAGE are two transductive graph learning methods, we\nuse a mask to indicate training and test nodes. For all the\ncompared methods, we use the same early stop criterion.\n4.1 Implementation\nThe number of embedding aggregation head H is 2, the\ndimension of edge embedding pis set as 32, and the hidden\nlayer of node embedding qis set as 128 dimensions. f1is\na two-layer modal attribute encoding module, and f2is\ntwo-layer fully connected network. The batch size is set as\n32, triplet loss margin is 1.0in experiment, random walk\nlength is set as 300, and the probability of returning to\nthe starting point is 0.05. The\u03bbin Equation 7 is tuned\nwith cross validation. The validation set is randomly se-\nlected from the training set by 10%. When the validation\nset loss does not decrease within 50epochs, the training\nwill be stopped. For comparing methods, we adjust the\nhyper-parameters according to the original paper to ac-\nAuthorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 20,2021 at 08:36:01 UTC from IEEE Xplore.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "c7ae3fa8-e717-4a52-9b40-c15d32848a2b": {"__data__": {"id_": "c7ae3fa8-e717-4a52-9b40-c15d32848a2b", "embedding": null, "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using a heterogeneous multi-modal graph neural network for corporate relative valuation?\n2. How is the validation set randomly selected from the training set for cross-validation in this study?\n3. What is the significance of the \u03bbin Equation 7 in tuning the hyper-parameters for the heterogeneous multi-modal graph neural network in this study?", "prev_section_summary": "The section discusses the use of Heterogeneous Multi-Modal Graph Neural Network (HM2) for corporate relative valuation. The authors use three meta-paths for m2vec unsupervised graph node embedding learning method, and the same input features and sampled neighbors set for GraphSAGE and GAT. The validation set is used to tune hyperparameters for the HM2 model. The section provides details on the implementation of the model, including the number of embedding aggregation heads, the dimension of edge and node embeddings, and the batch size. The authors also mention the use of cross-validation to tune the hyperparameters and the early stop criterion for comparing methods.", "section_summary": "The section discusses the use of a heterogeneous multi-modal graph neural network for corporate relative valuation. The purpose of this approach is to improve the accuracy of valuation by incorporating multiple types of data, such as financial and non-financial information. The validation set is randomly selected from the training set for cross-validation, and the \u03bbin Equation 7 is used to tune the hyper-parameters. The section also outlines the methodology for comparing different methods and the stopping criteria for training.", "excerpt_keywords": "1. AI assistant, 2. Care, 3. Respect, 4. Truth, 5. Utility, 6. Security, 7. Harmful content, 8. Ethical, 9. Prejudiced, 10. Positivity."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "8500b2d6-015e-436e-a558-aa4efc073e92", "node_type": "4", "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "9a57af184d633ae88d28a9adf02309b9bec7408e52023bc5e2801fb60d74ef6e"}, "2": {"node_id": "f8b751bb-40f9-4123-b6fc-5575dc5f2383", "node_type": "1", "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "1bbf85a6280a445b8b5e3bf5c1904529d0be184df24047edc1725801ef329848"}}, "hash": "286c256e8b51d6ef7260935ec604419d15f0772ed059c55e83119c162de7233a", "text": "is set as 300, and the probability of returning to\nthe starting point is 0.05. The\u03bbin Equation 7 is tuned\nwith cross validation. The validation set is randomly se-\nlected from the training set by 10%. When the validation\nset loss does not decrease within 50epochs, the training\nwill be stopped. For comparing methods, we adjust the\nhyper-parameters according to the original paper to ac-\nAuthorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 20,2021 at 08:36:01 UTC from IEEE Xplore.  Restrictions apply.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "ee7c1408-a55f-44ee-ac96-0fb40bbda6b8": {"__data__": {"id_": "ee7c1408-a55f-44ee-ac96-0fb40bbda6b8", "embedding": null, "metadata": {"page_label": "9", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the article \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\"?\n2. How does the proposed method, HM2, compare to traditional linear methods in terms of performance for corporate relative valuation?\n3. What are the evaluation metrics used to assess the performance of the proposed method, HM2, in the corporate relative valuation task?", "prev_section_summary": "The section discusses the use of a heterogeneous multi-modal graph neural network for corporate relative valuation. The purpose of this approach is to improve the accuracy of valuation by incorporating multiple types of data, such as financial and non-financial information. The validation set is randomly selected from the training set for cross-validation, and the \u03bbin Equation 7 is used to tune the hyper-parameters. The section also outlines the methodology for comparing different methods and the stopping criteria for training.", "section_summary": "The section discusses a research article titled \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\" by Yang et al. The purpose of the article is to propose a method, HM2, for corporate relative valuation using a heterogeneous multi-modal graph neural network. The proposed method is compared to traditional linear methods in terms of performance, and evaluation metrics such as Accuracy, Recall, Precision, and F1-measure are used to assess the performance of HM2. The results show that graph embedding methods outperform traditional linear methods and most methods achieve good performance in the corporate relative valuation task.", "excerpt_keywords": "1. HM2,\n2. Graph embedding,\n3. Corporate relative valuation,\n4. Node classification,\n5. Accuracy,\n6. Recall,\n7. Precision,\n8. F1-measure,\n9. Data imbalance,\n10. Nvidia 2080ti."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "c69dbb5a-e5cd-4e5b-baec-971425360fe2", "node_type": "4", "metadata": {"page_label": "9", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "7974425cbb776830c95e0300eafc863c309fde76b1a9e1f9dd231e7006b276f9"}, "3": {"node_id": "7f8e09db-ab4a-4a4e-8ecf-b23c92a915c9", "node_type": "1", "metadata": {"page_label": "9", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "8f679c98aec353d1566cb8ebd24a71adb7299ecc1fe9de77c735e60f5a34a50c"}}, "hash": "7f667442b5ce6125dac97f622193515dd1bd2ff6ae418a0558441a489bafdb69", "text": "1041-4347 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TKDE.2021.3080293, IEEE\nTransactions on Knowledge and Data Engineering\n9\nquire their optimal results. We implement HM2on a server\nwith GPU machines (Nvidia 2080ti). Code is available at:\nhttps://github.com/data-ming-and-application/HMM.\n4.2 Corporate Relative Valuation\nTo answer Q1, we design experiments to evaluate HM2on\ncorporate relative valuation task. We give both approximate\n(coarse-grained range y) and accurate (\ufb01ne-grained \u02c6y) valu-\nations, in which \u02c6ypdenotes the real value with logoperator\nof corporate valuation.\nSimilar to traditional node classi\ufb01cation task, we \ufb01rst use\ntraining data to build the model, then employ the learned\nmodel to predict nodes in test data. The ratio of training data\nis set to 10%, 30%, 50% and70%, and the remaining nodes\nare used for test. As a multi-class classi\ufb01cation problem,\nwe use Accuracy (Acc), Recall (Rec), Precision (Pre), and\nF1-measure as the evaluation metric. Note that we use\nweighted average measures of Rec, Pre and F1 considering\ndata imbalance problem. In addition, duplicated companies\nare removed from the experiments. Table 3 reports results\nof HM2and comparison methods. The results reveal that:\n1) Graph embedding methods are superior than traditional\nlinear method, which considers only the information of\ninput node itself; 2) Most methods achieve good perfor-\nmance in the corporate relative valuation, which", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "7f8e09db-ab4a-4a4e-8ecf-b23c92a915c9": {"__data__": {"id_": "7f8e09db-ab4a-4a4e-8ecf-b23c92a915c9", "embedding": null, "metadata": {"page_label": "9", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using Accuracy (Acc), Recall (Rec), Precision (Pre), and F1-measure as evaluation metrics in the corporate relative valuation task?\n2. How does the use of weighted average measures of Rec, Pre and F1 help address the data imbalance problem in the corporate relative valuation task?\n3. What is the significance of removing duplicated companies from the experiments in the corporate relative valuation task?", "prev_section_summary": "The section discusses a research article titled \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\" by Yang et al. The purpose of the article is to propose a method, HM2, for corporate relative valuation using a heterogeneous multi-modal graph neural network. The proposed method is compared to traditional linear methods in terms of performance, and evaluation metrics such as Accuracy, Recall, Precision, and F1-measure are used to assess the performance of HM2. The results show that graph embedding methods outperform traditional linear methods and most methods achieve good performance in the corporate relative valuation task.", "section_summary": "The section discusses the use of Accuracy (Acc), Recall (Rec), Precision (Pre), and F1-measure as evaluation metrics in the corporate relative valuation task, and how weighted average measures of Rec, Pre and F1 help address the data imbalance problem. The significance of removing duplicated companies from the experiments is also highlighted. The section compares the performance of graph embedding methods, including Heterogeneous Multi-Modal Graph Neural Network (HM2), with traditional linear methods and other attention-based graph models. The results show that HM2 achieves the best or comparable performance to comparison methods, and that it takes the relations as feature vectors and utilizes an extra mapping function for better learning the similarity between two nodes. The section also discusses the importance of considering heterogeneous neighbors and linkages comprehensively, and the effectiveness of the attention mechanism in employing the linkages. Finally, the section notes that HM2 employs the triplet loss by considering the embedding structure and can better reflect the global structure of graphs with the increase of training data.", "excerpt_keywords": "1. Graph embedding, 2. Corporate relative valuation, 3. Machine learning, 4. Domain experts, 5. Heterogeneous node attributes, 6. Linkages, 7. Attention mechanism, 8. LSTM, 9. Triplet loss, 10. Global structure of graphs."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "c69dbb5a-e5cd-4e5b-baec-971425360fe2", "node_type": "4", "metadata": {"page_label": "9", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "7974425cbb776830c95e0300eafc863c309fde76b1a9e1f9dd231e7006b276f9"}, "2": {"node_id": "ee7c1408-a55f-44ee-ac96-0fb40bbda6b8", "node_type": "1", "metadata": {"page_label": "9", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "7f667442b5ce6125dac97f622193515dd1bd2ff6ae418a0558441a489bafdb69"}, "3": {"node_id": "dbaf40b4-9063-4dd0-8a24-da12f3dc98cd", "node_type": "1", "metadata": {"page_label": "9", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "f3bedd53394428e3e657ad817bb73234f06880960d6be8ba09263985910bcc27"}}, "hash": "8f679c98aec353d1566cb8ebd24a71adb7299ecc1fe9de77c735e60f5a34a50c", "text": "problem,\nwe use Accuracy (Acc), Recall (Rec), Precision (Pre), and\nF1-measure as the evaluation metric. Note that we use\nweighted average measures of Rec, Pre and F1 considering\ndata imbalance problem. In addition, duplicated companies\nare removed from the experiments. Table 3 reports results\nof HM2and comparison methods. The results reveal that:\n1) Graph embedding methods are superior than traditional\nlinear method, which considers only the information of\ninput node itself; 2) Most methods achieve good perfor-\nmance in the corporate relative valuation, which re\ufb02ects\nthe effectiveness of machine learning models on simulating\nthe judgment of domain experts; 3) HM2achieves the best\nor comparable performance to comparison methods, which\nshows that HM2can encode effective node embedding\nfor valuation task by considering the heterogeneous node\nattributes and linkages comprehensively; 4) HM2performs\nbetter than another attention based graph model GAT, be-\ncause HM2considers heterogeneous neighbors and link-\nages comprehensively, and uses a more effective multi-\nhead attention mechanism that validates the effectiveness\nof heterogeneous neighbor construction and fusion; 5) HM2\nperforms better than HetGNN, which adopts LSTM to ag-\ngregate heterogeneous neighbors, and this indicates that\nattention mechanism can better employ the linkages; 6)\nHM2performs better than FAME, the reason is that HM2\ntakes the relations as feature vectors, and utilizes an extra\nmapping function \u03a8(\u00b7) for better learning the similarity\nbetween two nodes; and 7) With the increase of training\ndata, the performance of HM2improves faster than other\nmethods, for the reason that HM2employs the triplet loss by\nconsidering the embedding structure, and can better re\ufb02ect\nthe global structure of graphs with the increase of training\ndata. Moreover, we regard the corporate relative valuation\nas a regression problem, i.e.,", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "dbaf40b4-9063-4dd0-8a24-da12f3dc98cd": {"__data__": {"id_": "dbaf40b4-9063-4dd0-8a24-da12f3dc98cd", "embedding": null, "metadata": {"page_label": "9", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the difference between Heterogeneous Multi-Modal Graph Neural Network (HM2) and FAME in terms of their performance on corporate relative valuation prediction?\n2. How does the use of triplet loss in HM2 improve its performance compared to other methods in predicting corporate relative valuation?\n3. What are the limitations of using deep graph models for corporate relative valuation prediction, and how do they affect the performance of these methods?", "prev_section_summary": "The section discusses the use of Accuracy (Acc), Recall (Rec), Precision (Pre), and F1-measure as evaluation metrics in the corporate relative valuation task, and how weighted average measures of Rec, Pre and F1 help address the data imbalance problem. The significance of removing duplicated companies from the experiments is also highlighted. The section compares the performance of graph embedding methods, including Heterogeneous Multi-Modal Graph Neural Network (HM2), with traditional linear methods and other attention-based graph models. The results show that HM2 achieves the best or comparable performance to comparison methods, and that it takes the relations as feature vectors and utilizes an extra mapping function for better learning the similarity between two nodes. The section also discusses the importance of considering heterogeneous neighbors and linkages comprehensively, and the effectiveness of the attention mechanism in employing the linkages. Finally, the section notes that HM2 employs the triplet loss by considering the embedding structure and can better reflect the global structure of graphs with the increase of training data.", "section_summary": "The section discusses the use of Heterogeneous Multi-Modal Graph Neural Network (HM2) for predicting corporate relative valuation. HM2 takes the relations as feature vectors and utilizes an extra mapping function for better learning the similarity between two nodes. The use of triplet loss in HM2 improves its performance compared to other methods in predicting corporate relative valuation. The section also discusses the limitations of using deep graph models for corporate relative valuation prediction and how they affect the performance of these methods. The results reveal that HM2 achieves the best or comparable performance, even under low training data ratio, and promotes between deep methods and linear method are not significant. The section also mentions the data privacy and field limitations, as well as the feature missing and information insufficiency, which affect the training of deep models.", "excerpt_keywords": "1. HM2, 2. graph embedding, 3. relative valuation, 4. machine learning, 5. deep learning, 6. feature vectors, 7. triplet loss, 8. data privacy, 9. field limitations, 10. information insufficiency"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "c69dbb5a-e5cd-4e5b-baec-971425360fe2", "node_type": "4", "metadata": {"page_label": "9", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "7974425cbb776830c95e0300eafc863c309fde76b1a9e1f9dd231e7006b276f9"}, "2": {"node_id": "7f8e09db-ab4a-4a4e-8ecf-b23c92a915c9", "node_type": "1", "metadata": {"page_label": "9", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "8f679c98aec353d1566cb8ebd24a71adb7299ecc1fe9de77c735e60f5a34a50c"}, "3": {"node_id": "faf2efec-9e29-4db3-8c8a-ec87e4f3548e", "node_type": "1", "metadata": {"page_label": "9", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "2d23682d97f5fc5269aef7ba551836af9349aa24144002b273cfaed70066daa5"}}, "hash": "f3bedd53394428e3e657ad817bb73234f06880960d6be8ba09263985910bcc27", "text": "than FAME, the reason is that HM2\ntakes the relations as feature vectors, and utilizes an extra\nmapping function \u03a8(\u00b7) for better learning the similarity\nbetween two nodes; and 7) With the increase of training\ndata, the performance of HM2improves faster than other\nmethods, for the reason that HM2employs the triplet loss by\nconsidering the embedding structure, and can better re\ufb02ect\nthe global structure of graphs with the increase of training\ndata. Moreover, we regard the corporate relative valuation\nas a regression problem, i.e., \u2113m=\u2225\u02c6yp\ni\u2212\u03b8\u22a4f2(v)\u22252. Table 4\nreports the MSE results of HM2and comparison methods.\nThe results reveal that: 1) The machine learning methods\nalso have considerable performance on accurate relative\nvaluation prediction; and 2) HM2also achieves the best\nor comparable performance, which is much better than\ncomparison methods even under low training data ratio,\ni.e., HM2performs better with only 10% training data.\nA notable phenomenon is that various methods do not\nhave signi\ufb01cant performance: 1) Even the performance of\nthe best method are not signi\ufb01cant; and 2) The promotions\nbetween deep methods and linear method, and the promo-\ntions between HM2and other deep graph models are notsigni\ufb01cant. This is because: 1) Considering the data privacy\nand \ufb01eld limitations, the amount of data is relatively small,\nwhich affects the training of deep models; and 2) Consid-\nering the feature missing and information insuf\ufb01ciency, the\ninformation contained in raw multi-modal data is limited.\n5 10 15 20 25 30 35 400.4100.4150.4200.4250.4300.4350.4400.4450.450\nAccuracy\nF1\n(a) Accuracy &F1\n5 10 15 20 25 30", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "faf2efec-9e29-4db3-8c8a-ec87e4f3548e": {"__data__": {"id_": "faf2efec-9e29-4db3-8c8a-ec87e4f3548e", "embedding": null, "metadata": {"page_label": "9", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the ablation studies conducted in the paper \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\"?\n2. How does the performance of the Heterogeneous Multi-Modal Graph Neural Network (HM2) vary with the size of the sampled neighbors?\n3. What is the role of the relation embedding in the Heterogeneous Multi-Modal Graph Neural Network (HM2) and how does its removal affect the performance of the model?", "prev_section_summary": "The section discusses the use of Heterogeneous Multi-Modal Graph Neural Network (HM2) for predicting corporate relative valuation. HM2 takes the relations as feature vectors and utilizes an extra mapping function for better learning the similarity between two nodes. The use of triplet loss in HM2 improves its performance compared to other methods in predicting corporate relative valuation. The section also discusses the limitations of using deep graph models for corporate relative valuation prediction and how they affect the performance of these methods. The results reveal that HM2 achieves the best or comparable performance, even under low training data ratio, and promotes between deep methods and linear method are not significant. The section also mentions the data privacy and field limitations, as well as the feature missing and information insufficiency, which affect the training of deep models.", "section_summary": "The section discusses the purpose of ablation studies conducted in the paper \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\". The ablation studies were designed to evaluate the performance of the Heterogeneous Multi-Modal Graph Neural Network (HM2) and its variants. The variants included HM2-N, HM2-L, HM2-FC, HM2-R, HM2-A, and HM2-B. The results of the prediction were analyzed to determine the role of each module in HM2 and how its removal affects the performance of the model. The section also discusses the influence of sampled neighbor size on the performance of HM2 and the role of relation embedding in the model.", "excerpt_keywords": "1. Heterogeneous Multi-Modal Network,\n2. Ablation Study,\n3. Direct Neighbors,\n4. Transformer Based Attribute Encoding,\n5. Bi-LSTM,\n6. Fully Connected Network,\n7. Relation Embedding,\n8. Attention Based Multi-Modal Aggregation,\n9. Triplet Loss,\n10. Performance Measure."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "c69dbb5a-e5cd-4e5b-baec-971425360fe2", "node_type": "4", "metadata": {"page_label": "9", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "7974425cbb776830c95e0300eafc863c309fde76b1a9e1f9dd231e7006b276f9"}, "2": {"node_id": "dbaf40b4-9063-4dd0-8a24-da12f3dc98cd", "node_type": "1", "metadata": {"page_label": "9", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "f3bedd53394428e3e657ad817bb73234f06880960d6be8ba09263985910bcc27"}, "3": {"node_id": "4b8db73f-108b-441c-be29-3db8f6bc5d78", "node_type": "1", "metadata": {"page_label": "9", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "cc9275ec106841a0f3201ce691e9ce48bffcb38097fa1cb41fb6a5f0692fbc30"}}, "hash": "2d23682d97f5fc5269aef7ba551836af9349aa24144002b273cfaed70066daa5", "text": "This is because: 1) Considering the data privacy\nand \ufb01eld limitations, the amount of data is relatively small,\nwhich affects the training of deep models; and 2) Consid-\nering the feature missing and information insuf\ufb01ciency, the\ninformation contained in raw multi-modal data is limited.\n5 10 15 20 25 30 35 400.4100.4150.4200.4250.4300.4350.4400.4450.450\nAccuracy\nF1\n(a) Accuracy &F1\n5 10 15 20 25 30 352.9252.9502.9753.0003.0253.0503.0753.100\nMSE\n(b) MSE\nFig. 5. In\ufb02uence of sampled neighbor size, x-axis denotes the neighbor\nsize and y-axis represents performance measure.\n4.3 Analysis of HM2\nTo answer Q2, we design ablation studies for evaluation.\nAblation Study\nTo explore the role of each module in HM2, we conduct\nextra ablation studies to evaluate performances of several\nvariants, including:\n\u2022HM2-N: The variant of HM2that only adopts the\ndirect neighbors, without considering the higher-order\nneighbors;\n\u2022HM2-L: The variant of HM2that replaces the trans-\nformer based attribute encoding module with tradi-\ntional Bi-LSTM to encode heterogeneous node;\n\u2022HM2-FC: The variant of HM2that replaces the trans-\nformer based attribute encoding module with fully\nconnected network to encode heterogeneous node;\n\u2022HM2-R: The variant of HM2that doesn\u2019t consider the\nrelation embedding in Eq. 3;\n\u2022HM2-A: The variant of HM2that replaces attention\nbased multi-modal aggregation module with directly\nconcatenating multi-modal embeddings;\n\u2022HM2-B: The variant of HM2that removes the triplet\nloss in Eq. 7.\nThe results of prediction are", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "4b8db73f-108b-441c-be29-3db8f6bc5d78": {"__data__": {"id_": "4b8db73f-108b-441c-be29-3db8f6bc5d78", "embedding": null, "metadata": {"page_label": "9", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the study and what problem does it aim to solve?\n2. What are the different variants of the Heterogeneous Multi-Modal Graph Neural Network (HM2) and how do they differ from each other?\n3. What are the results of the study and how do they compare to other methods for corporate relative valuation?", "prev_section_summary": "The section discusses the purpose of ablation studies conducted in the paper \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\". The ablation studies were designed to evaluate the performance of the Heterogeneous Multi-Modal Graph Neural Network (HM2) and its variants. The variants included HM2-N, HM2-L, HM2-FC, HM2-R, HM2-A, and HM2-B. The results of the prediction were analyzed to determine the role of each module in HM2 and how its removal affects the performance of the model. The section also discusses the influence of sampled neighbor size on the performance of HM2 and the role of relation embedding in the model.", "section_summary": "The section discusses a study that aims to solve the problem of corporate relative valuation using a Heterogeneous Multi-Modal Graph Neural Network (HM2). The study compares the performance of different variants of HM2, including HM2-R, HM2-A, and HM2-B, to other methods for corporate relative valuation. The results reveal that HM2 outperforms other methods, demonstrating the effectiveness of neighbor sampling and linkage-aware multi-head attention based encoding. The study is reported in tables 5 and 6, and the variant of HM2 that replaces the trans-former based attribute encoding module with a fully connected network to encode heterogeneous nodes is particularly effective.", "excerpt_keywords": "1. Heterogeneous Multi-Modal Networks\n2. Attention-based Multi-Modal Aggregation\n3. Transformer-based Attribute Encoding\n4. Linkage-aware Multi-Head Attention\n5. Neighbor Sampling\n6. Multi-Modal Embedding Generalization\n7. Triplet Loss\n8. Fully Connected Network\n9. HM2-R, HM2-A, HM2-B\n10. Prediction Results"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "c69dbb5a-e5cd-4e5b-baec-971425360fe2", "node_type": "4", "metadata": {"page_label": "9", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "7974425cbb776830c95e0300eafc863c309fde76b1a9e1f9dd231e7006b276f9"}, "2": {"node_id": "faf2efec-9e29-4db3-8c8a-ec87e4f3548e", "node_type": "1", "metadata": {"page_label": "9", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "2d23682d97f5fc5269aef7ba551836af9349aa24144002b273cfaed70066daa5"}}, "hash": "cc9275ec106841a0f3201ce691e9ce48bffcb38097fa1cb41fb6a5f0692fbc30", "text": "The variant of HM2that replaces the trans-\nformer based attribute encoding module with fully\nconnected network to encode heterogeneous node;\n\u2022HM2-R: The variant of HM2that doesn\u2019t consider the\nrelation embedding in Eq. 3;\n\u2022HM2-A: The variant of HM2that replaces attention\nbased multi-modal aggregation module with directly\nconcatenating multi-modal embeddings;\n\u2022HM2-B: The variant of HM2that removes the triplet\nloss in Eq. 7.\nThe results of prediction are reported in Table 5 and\nTable 6. They reveal that: 1) HM2behaves better than HM2-\nN, which demonstrates that neighbor sampling is effective\nfor subsequent operation and embedding generalization;\n2) HM2behaves better than HM2-L and HM2-FC, which\nshows that linkage-aware multi-head attention based en-\ncoding outperforms other methods without considering the\nAuthorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 20,2021 at 08:36:01 UTC from IEEE Xplore.  Restrictions apply.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "ba8f4e68-acd7-4a47-a776-80b310febb8f": {"__data__": {"id_": "ba8f4e68-acd7-4a47-a776-80b310febb8f", "embedding": null, "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the performance of the Heterogeneous Multi-Modal Graph Neural Network (HM2-N) model in predicting corporate relative valuation with different training data ratios?\n2. How does the precision of the HM2-N model vary with different training data ratios?\n3. What is the impact of the heterogeneity of the multi-modal graph on the performance of the HM2-N model in predicting corporate relative valuation?", "prev_section_summary": "The section discusses a study that aims to solve the problem of corporate relative valuation using a Heterogeneous Multi-Modal Graph Neural Network (HM2). The study compares the performance of different variants of HM2, including HM2-R, HM2-A, and HM2-B, to other methods for corporate relative valuation. The results reveal that HM2 outperforms other methods, demonstrating the effectiveness of neighbor sampling and linkage-aware multi-head attention based encoding. The study is reported in tables 5 and 6, and the variant of HM2 that replaces the trans-former based attribute encoding module with a fully connected network to encode heterogeneous nodes is particularly effective.", "section_summary": "The section discusses the performance of a Heterogeneous Multi-Modal Graph Neural Network (HM2-N) model in predicting corporate relative valuation with different training data ratios. The study compares the accuracy and precision of the HM2-N model with other models (HM2-L, HM2-FC, HM2-A, HM2-B, HM2-R, and HM2) using an ablation study. The results show that the HM2-N model outperforms the other models in terms of accuracy and precision, with the best results achieved when using a training data ratio of 70%. The heterogeneity of the multi-modal graph has a significant impact on the performance of the HM2-N model, with better results achieved when the graph is more heterogeneous.", "excerpt_keywords": "1. Ablation study\n2. HM2-N\n3. HM2-L\n4. HM2-FC\n5. HM2-A\n6. HM2-B\n7. HM2-R\n8. HM2\n9. Accuracy\n10. Precision"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "721e6fe2-2be3-4029-8b16-c9648c3e2621", "node_type": "4", "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "f48079b147206e24f1a603002ca88b8dd9ee1cedd13676fe17fa8d88d312fecf"}, "3": {"node_id": "ef19d03d-3442-458c-adef-cb8a162d42bb", "node_type": "1", "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "1af89e7068bbd3ab2280f2b480dc3224b549213e5ab812dc2b207d3c53294901"}}, "hash": "1e51c6f880b4f6cdc099ac1e55bb9f30eb6b328adc9ddf444f06a2c90de1f4eb", "text": "1041-4347 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TKDE.2021.3080293, IEEE\nTransactions on Knowledge and Data Engineering\n10\nTABLE 5\nAblation Study (y ), percentage denotes training data ratio. The best results are highlighted in bold.\nMetric HM2-N HM2-L HM2-FC HM2-A HM2-B HM2-R HM2\nAccuracy10% 0.335\u00b10.003 0.329\u00b10.006 0.334\u00b10.009 0.341\u00b10.003 0.326\u00b10.006 0.334\u00b10.005 0.346\u00b10.008\n30% 0.366\u00b10.003 0.380\u00b10.006 0.347\u00b10.011 0.388\u00b10.002 0.380\u00b10.004 0.385\u00b10.005 0.388\u00b10.004\n50% 0.396\u00b10.004 0.389\u00b10.014 0.381\u00b10.016 0.381\u00b10.006 0.377\u00b10.008 0.396\u00b10.008 0.410\u00b10.008\n70% 0.427\u00b10.009 0.434\u00b10.026 0.420\u00b10.017 0.430\u00b10.011 0.424\u00b10.008 0.433\u00b10.006 0.446\u00b10.007\nPrecision10% 0.333\u00b10.003 0.329\u00b10.031 0.337\u00b10.008 0.340\u00b10.002 0.337\u00b10.007 0.334\u00b10.005 0.351\u00b10.006\n30% 0.379\u00b10.005 0.375\u00b10.003", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "ef19d03d-3442-458c-adef-cb8a162d42bb": {"__data__": {"id_": "ef19d03d-3442-458c-adef-cb8a162d42bb", "embedding": null, "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the performance of the heterogeneous multi-modal graph neural network in predicting corporate relative valuation with respect to different evaluation metrics such as precision, recall, and F1 score?\n2. How does the heterogeneous multi-modal graph neural network compare with other state-of-the-art models in terms of accuracy and efficiency in predicting corporate relative valuation?\n3. What are the key features and characteristics of the heterogeneous multi-modal graph neural network that contribute to its superior performance in predicting corporate relative valuation compared to other models?", "prev_section_summary": "The section discusses the performance of a Heterogeneous Multi-Modal Graph Neural Network (HM2-N) model in predicting corporate relative valuation with different training data ratios. The study compares the accuracy and precision of the HM2-N model with other models (HM2-L, HM2-FC, HM2-A, HM2-B, HM2-R, and HM2) using an ablation study. The results show that the HM2-N model outperforms the other models in terms of accuracy and precision, with the best results achieved when using a training data ratio of 70%. The heterogeneity of the multi-modal graph has a significant impact on the performance of the HM2-N model, with better results achieved when the graph is more heterogeneous.", "section_summary": "The section discusses the performance of a heterogeneous multi-modal graph neural network in predicting corporate relative valuation. The network is compared with other state-of-the-art models in terms of accuracy and efficiency. The key features and characteristics of the network that contribute to its superior performance are also discussed. The evaluation metrics used to assess the performance of the network include precision, recall, and F1 score. The section provides tables with the performance of the network on these metrics for different evaluation thresholds.", "excerpt_keywords": "1. Assist,\n2. Care,\n3. Respect,\n4. Truth,\n5. Utility,\n6. Security,\n7. Harmful,\n8. Unethical,\n9. Prejudiced,\n10. Negative."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "721e6fe2-2be3-4029-8b16-c9648c3e2621", "node_type": "4", "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "f48079b147206e24f1a603002ca88b8dd9ee1cedd13676fe17fa8d88d312fecf"}, "2": {"node_id": "ba8f4e68-acd7-4a47-a776-80b310febb8f", "node_type": "1", "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "1e51c6f880b4f6cdc099ac1e55bb9f30eb6b328adc9ddf444f06a2c90de1f4eb"}, "3": {"node_id": "14a82aaf-27b8-4cf4-b149-7e5d5eceb628", "node_type": "1", "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "23ad7311f098bef6763b40a8a9da8fa3fd625aa362900bb3afa32365204072bb"}}, "hash": "1af89e7068bbd3ab2280f2b480dc3224b549213e5ab812dc2b207d3c53294901", "text": "0.427\u00b10.009 0.434\u00b10.026 0.420\u00b10.017 0.430\u00b10.011 0.424\u00b10.008 0.433\u00b10.006 0.446\u00b10.007\nPrecision10% 0.333\u00b10.003 0.329\u00b10.031 0.337\u00b10.008 0.340\u00b10.002 0.337\u00b10.007 0.334\u00b10.005 0.351\u00b10.006\n30% 0.379\u00b10.005 0.375\u00b10.003 0.372\u00b10.012 0.393\u00b10.004 0.379\u00b10.004 0.388\u00b10.005 0.395\u00b10.004\n50% 0.385\u00b10.006 0.377\u00b10.016 0.382\u00b10.026 0.374\u00b10.005 0.377\u00b10.008 0.390\u00b10.016 0.405\u00b10.012\n70% 0.423\u00b10.008 0.421\u00b10.034 0.418\u00b10.023 0.432\u00b10.010 0.428\u00b10.009 0.425\u00b10.007 0.428\u00b10.004\nRecall10% 0.335\u00b10.003 0.329\u00b10.006 0.334\u00b10.009 0.341\u00b10.002 0.326\u00b10.010 0.334\u00b10.005 0.346\u00b10.008\n30% 0.366\u00b10.003 0.380\u00b10.006 0.347\u00b10.011 0.388\u00b10.004 0.380\u00b10.004 0.385\u00b10.005 0.388\u00b10.004\n50% 0.396\u00b10.004 0.389\u00b10.014 0.381\u00b10.016 0.381\u00b10.009 0.377\u00b10.008 0.396\u00b10.008 0.410\u00b10.012\n70% 0.427\u00b10.009 0.434\u00b10.025 0.420\u00b10.017 0.430\u00b10.013", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "14a82aaf-27b8-4cf4-b149-7e5d5eceb628": {"__data__": {"id_": "14a82aaf-27b8-4cf4-b149-7e5d5eceb628", "embedding": null, "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the study \"Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network\"?\n2. What is the performance of the proposed model in terms of MSE10%, F1-measure10%, and F1-measure30% for different training data ratios?\n3. How does the ablation study of the proposed model affect the performance of the model in terms of MSE10%, F1-measure10%, and F1-measure30% for different training data ratios?", "prev_section_summary": "The section discusses the performance of a heterogeneous multi-modal graph neural network in predicting corporate relative valuation. The network is compared with other state-of-the-art models in terms of accuracy and efficiency. The key features and characteristics of the network that contribute to its superior performance are also discussed. The evaluation metrics used to assess the performance of the network include precision, recall, and F1 score. The section provides tables with the performance of the network on these metrics for different evaluation thresholds.", "section_summary": "The section discusses the study \"Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network\" by Yang et al. The purpose of the study is to develop a model for corporate relative valuation using a heterogeneous multi-modal graph neural network. The performance of the proposed model is evaluated in terms of MSE10%, F1-measure10%, and F1-measure30% for different training data ratios. An ablation study is also conducted to investigate the impact of different components of the model on its performance. The results show that the proposed model outperforms the baseline models in terms of MSE10% and F1-measure10%, and that the ablation study has a significant impact on the performance of the model.", "excerpt_keywords": "1. HM2-N, HM2-L, HM2-FC, HM2-A, HM2-B, HM2-R, MSE10%, ablation study, training data ratio, fairness, positivity, utility, care, respect, truth, ethical, unethical, prejudiced, negative, F1-measure, 0.331, 0.327, 0.334, 0.340, 0.329, 0.333, 0.356, 0.371, 0.341, 0.387, 0.376, 0.382, 0.376, 0.387, 0.357, 0.368, 0.347, 0.376, 0.357, 0.387, 0.400, 0.423, 0.424, 0.425, 0.427"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "721e6fe2-2be3-4029-8b16-c9648c3e2621", "node_type": "4", "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "f48079b147206e24f1a603002ca88b8dd9ee1cedd13676fe17fa8d88d312fecf"}, "2": {"node_id": "ef19d03d-3442-458c-adef-cb8a162d42bb", "node_type": "1", "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "1af89e7068bbd3ab2280f2b480dc3224b549213e5ab812dc2b207d3c53294901"}, "3": {"node_id": "228042aa-b36d-4bfe-a3a7-e4c838cb1a8d", "node_type": "1", "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "8ec9692e1ced98ad827c1fb059457fd838cca7c7da8c0627bb09dcf726b028fe"}}, "hash": "23ad7311f098bef6763b40a8a9da8fa3fd625aa362900bb3afa32365204072bb", "text": "0.380\u00b10.006 0.347\u00b10.011 0.388\u00b10.004 0.380\u00b10.004 0.385\u00b10.005 0.388\u00b10.004\n50% 0.396\u00b10.004 0.389\u00b10.014 0.381\u00b10.016 0.381\u00b10.009 0.377\u00b10.008 0.396\u00b10.008 0.410\u00b10.012\n70% 0.427\u00b10.009 0.434\u00b10.025 0.420\u00b10.017 0.430\u00b10.013 0.424\u00b10.008 0.433\u00b10.006 0.446\u00b10.005\nF1-measure10% 0.331\u00b10.003 0.327\u00b10.022 0.334\u00b10.010 0.340\u00b10.003 0.329\u00b10.006 0.333\u00b10.004 0.340\u00b10.009\n30% 0.356\u00b10.004 0.371\u00b10.006 0.341\u00b10.013 0.387\u00b10.002 0.376\u00b10.004 0.382\u00b10.006 0.376\u00b10.005\n50% 0.388\u00b10.004 0.368\u00b10.028 0.347\u00b10.024 0.376\u00b10.006 0.357\u00b10.009 0.387\u00b10.013 0.400\u00b10.008\n70% 0.418\u00b10.010 0.423\u00b10.026 0.418\u00b10.023 0.425\u00b10.011 0.423\u00b10.007 0.423\u00b10.006 0.424\u00b10.007\nTABLE 6\nAblation Study ( \u02c6y), percentage denotes training data ratio. The best results are highlighted in bold.\nMetric HM2-N HM2-L HM2-FC HM2-A HM2-B HM2-R HM2\nMSE10% 5.223\u00b10.231 3.921\u00b10.559 5.304\u00b10.487 4.091\u00b10.658", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "228042aa-b36d-4bfe-a3a7-e4c838cb1a8d": {"__data__": {"id_": "228042aa-b36d-4bfe-a3a7-e4c838cb1a8d", "embedding": null, "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. How does the heterogeneous multi-modal graph neural network (HM2) perform in predicting corporate relative valuation compared to other methods?\n2. What is the role of linkages and self-attention mechanism in the effectiveness of HM2 in predicting corporate relative valuation?\n3. How does the triplet loss function contribute to the discriminative learning of node embeddings in HM2 for predicting corporate relative valuation?", "prev_section_summary": "The section discusses the study \"Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network\" by Yang et al. The purpose of the study is to develop a model for corporate relative valuation using a heterogeneous multi-modal graph neural network. The performance of the proposed model is evaluated in terms of MSE10%, F1-measure10%, and F1-measure30% for different training data ratios. An ablation study is also conducted to investigate the impact of different components of the model on its performance. The results show that the proposed model outperforms the baseline models in terms of MSE10% and F1-measure10%, and that the ablation study has a significant impact on the performance of the model.", "section_summary": "The section discusses the performance of a heterogeneous multi-modal graph neural network (HM2) in predicting corporate relative valuation compared to other methods. The section also examines the role of linkages and self-attention mechanism in the effectiveness of HM2 in predicting corporate relative valuation. Additionally, the section explores how the triplet loss function contributes to the discriminative learning of node embeddings in HM2 for predicting corporate relative valuation. The section presents results from an ablation study that compares the performance of HM2 with different variations of the model, including HM2-N, HM2-L, HM2-FC, HM2-A, HM2-B, HM2-R, and HM2. The section concludes by summarizing the key findings and discussing the implications of the results.", "excerpt_keywords": "1. Ablation study, 2. HM2-N, HM2-L, HM2-FC, HM2-A, HM2-B, HM2-R, HM2, 3. MSE10%, 30%, 50%, 70%, 4. Linkages, 5. Self-attention mechanism, 6. Node embedding, 7. Triplet loss, 8. Graph structure, 9. Discriminative, 10. Prediction."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "721e6fe2-2be3-4029-8b16-c9648c3e2621", "node_type": "4", "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "f48079b147206e24f1a603002ca88b8dd9ee1cedd13676fe17fa8d88d312fecf"}, "2": {"node_id": "14a82aaf-27b8-4cf4-b149-7e5d5eceb628", "node_type": "1", "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "23ad7311f098bef6763b40a8a9da8fa3fd625aa362900bb3afa32365204072bb"}, "3": {"node_id": "33d77f1b-639c-42f9-8483-4026f5c09bef", "node_type": "1", "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "6b7c344e46f7343fa01fe820a04d8f5297ae165105197f1b083d64798f08c5d7"}}, "hash": "8ec9692e1ced98ad827c1fb059457fd838cca7c7da8c0627bb09dcf726b028fe", "text": "0.418\u00b10.023 0.425\u00b10.011 0.423\u00b10.007 0.423\u00b10.006 0.424\u00b10.007\nTABLE 6\nAblation Study ( \u02c6y), percentage denotes training data ratio. The best results are highlighted in bold.\nMetric HM2-N HM2-L HM2-FC HM2-A HM2-B HM2-R HM2\nMSE10% 5.223\u00b10.231 3.921\u00b10.559 5.304\u00b10.487 4.091\u00b10.658 5.808\u00b10.269 4.013\u00b1 0.216 3.919\u00b10.077\n30% 3.807\u00b10.104 4.027\u00b10.905 4.022\u00b10.151 3.801\u00b10.068 3.766\u00b10.086 3.605\u00b1 0.120 3.481\u00b10.051\n50% 3.731\u00b10.066 3.708\u00b10.261 3.800\u00b10.138 3.505\u00b10.034 3.895\u00b10.053 3.477\u00b1 0.076 3.432\u00b10.071\n70% 3.267\u00b10.065 3.167\u00b10.352 3.034\u00b10.162 3.114\u00b10.031 3.251\u00b10.068 3.014\u00b10.058 2.951\u00b10.084\nlinkages, and is bene\ufb01cial for learning attribute interactions;\n3) HM2behaves competitive to HM2-A, which reveals that\nthe self-attention mechanism has a slight advantage, and\nboth kinds of neighbors have relative contributes for predic-\ntion; 4) HM2performances better than HM2-R, which indi-\ncates the effectiveness of linkage representation in learning\nnode embedding; and 5) HM2performs better than HM2-B,\nwhich shows that triplet loss can take graph structure into\nfull account to learn more discriminative", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "33d77f1b-639c-42f9-8483-4026f5c09bef": {"__data__": {"id_": "33d77f1b-639c-42f9-8483-4026f5c09bef", "embedding": null, "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the study in the paper \"Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf\"?\n2. How does the self-attention mechanism contribute to the effectiveness of the Heterogeneous MultiModal Graph Neural Network in predicting corporate relative valuation?\n3. What is the significance of the linkage type set in the CRV graph and how does it affect the embedding learning of company nodes in the Heterogeneous MultiModal Graph Neural Network?", "prev_section_summary": "The section discusses the performance of a heterogeneous multi-modal graph neural network (HM2) in predicting corporate relative valuation compared to other methods. The section also examines the role of linkages and self-attention mechanism in the effectiveness of HM2 in predicting corporate relative valuation. Additionally, the section explores how the triplet loss function contributes to the discriminative learning of node embeddings in HM2 for predicting corporate relative valuation. The section presents results from an ablation study that compares the performance of HM2 with different variations of the model, including HM2-N, HM2-L, HM2-FC, HM2-A, HM2-B, HM2-R, and HM2. The section concludes by summarizing the key findings and discussing the implications of the results.", "section_summary": "The section discusses a study on using a Heterogeneous MultiModal Graph Neural Network (HM2) for predicting corporate relative valuation. The study focuses on the influence of linkages on embedding learning and conducts ablation studies to evaluate the relationships between different linkage types. The linkage type set of the CRV graph includes company-company, company-member, and member-member linkages. The study finds that the performance degradation of w/o m-m (without member-member linkages) is not obvious, while w/o c-m (without company-member linkages) and w/o e (without distinguishing edge types) result in significant performance degradation. Overall, the study suggests that the linkage type set plays an important role in learning attribute interactions and predicting corporate relative valuation.", "excerpt_keywords": "1. CRV graph\n2. company nodes\n3. member nodes\n4. linkage types\n5. random walk sampling\n6. ablation studies\n7. company-company linkages\n8. company-member linkages\n9. member-member linkages\n10. triplet loss"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "721e6fe2-2be3-4029-8b16-c9648c3e2621", "node_type": "4", "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "f48079b147206e24f1a603002ca88b8dd9ee1cedd13676fe17fa8d88d312fecf"}, "2": {"node_id": "228042aa-b36d-4bfe-a3a7-e4c838cb1a8d", "node_type": "1", "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "8ec9692e1ced98ad827c1fb059457fd838cca7c7da8c0627bb09dcf726b028fe"}, "3": {"node_id": "f5231109-65d5-43e0-8568-e77d8249a5ee", "node_type": "1", "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "0fef31a920157b4546829aa7e212da895c812ba0d3183927e74983f2cbb08f7b"}}, "hash": "6b7c344e46f7343fa01fe820a04d8f5297ae165105197f1b083d64798f08c5d7", "text": "and is bene\ufb01cial for learning attribute interactions;\n3) HM2behaves competitive to HM2-A, which reveals that\nthe self-attention mechanism has a slight advantage, and\nboth kinds of neighbors have relative contributes for predic-\ntion; 4) HM2performances better than HM2-R, which indi-\ncates the effectiveness of linkage representation in learning\nnode embedding; and 5) HM2performs better than HM2-B,\nwhich shows that triplet loss can take graph structure into\nfull account to learn more discriminative embedding.\n4.4 In\ufb02uence of Linkages\nIn detail, the linkage type set of the CRV graph consists:\n1)company-company linkages, 2) company-member linkages,\nand 3) member-member linkages. The main statistics are in\nTable 1 of the main body. Although HM2only focused on\nthe embedding learning of company nodes in this paper,\nwhen collecting company nodes\u2019 neighbors, HM2utilizes\nthe random walk sampling for neighbor construction of\neach node following [14]. Therefore, the company-member\nand member-member linkages can help collect high-order\nmember neighbors for each node.\nTo explore the in\ufb02uence of linkages on embedding\nlearning, we have conducted ablation studies to evaluate\nthe mentioned relationships: 1) w/o m-m: HM2without\nmember-member linkages, i.e., HM2that only samples the\ncompany\u2019s direct member neighbors; 2) w/o c-m: HM2\nwithout company-member linkages, i.e., HM2that does not\nsample company\u2019s member neighbors; 3) w/o e: HM2that\ndoes not distinguish edge types, i.e., the linkage is 1if two\nnodes connected, otherwise is 0.\nTable 7 records the results, and they reveal that: 1) the\nperformance degradation of w/o m-m is not obvious, itTABLE 7\nCorporate relative valuation prediction results with ablation", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "f5231109-65d5-43e0-8568-e77d8249a5ee": {"__data__": {"id_": "f5231109-65d5-43e0-8568-e77d8249a5ee", "embedding": null, "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the study and what problem does it aim to solve?\n2. What is the performance of the Heterogeneous Multi-Modal Graph Neural Network (HM2) model in predicting corporate relative valuation when considering different linkage settings?\n3. How does the inclusion of company-member linkages (c-m) and edge types (e) affect the performance of the HM2 model in predicting corporate relative valuation?", "prev_section_summary": "The section discusses a study on using a Heterogeneous MultiModal Graph Neural Network (HM2) for predicting corporate relative valuation. The study focuses on the influence of linkages on embedding learning and conducts ablation studies to evaluate the relationships between different linkage types. The linkage type set of the CRV graph includes company-company, company-member, and member-member linkages. The study finds that the performance degradation of w/o m-m (without member-member linkages) is not obvious, while w/o c-m (without company-member linkages) and w/o e (without distinguishing edge types) result in significant performance degradation. Overall, the study suggests that the linkage type set plays an important role in learning attribute interactions and predicting corporate relative valuation.", "section_summary": "The section discusses a study that aims to develop a Heterogeneous Multi-Modal Graph Neural Network (HM2) model for predicting corporate relative valuation. The study considers different linkage settings and analyzes the impact of company-member linkages (c-m) and edge types (e) on the performance of the HM2 model. The results reveal that the inclusion of c-m and e improves the accuracy, precision, and recall of the HM2 model in predicting corporate relative valuation.", "excerpt_keywords": "1. Corporate relative valuation prediction\n2. Ablation studies\n3. Linkage settings\n4. Performance degradation\n5. Accuracy\n6. Precision\n7. Recall\n8. Company-member linkages\n9. Edge types\n10. HM2"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "721e6fe2-2be3-4029-8b16-c9648c3e2621", "node_type": "4", "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "f48079b147206e24f1a603002ca88b8dd9ee1cedd13676fe17fa8d88d312fecf"}, "2": {"node_id": "33d77f1b-639c-42f9-8483-4026f5c09bef", "node_type": "1", "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "6b7c344e46f7343fa01fe820a04d8f5297ae165105197f1b083d64798f08c5d7"}, "3": {"node_id": "4628dd81-f2b7-4da3-8457-109259410331", "node_type": "1", "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "2645d0af1c3bbadf123eef3e450d6f901ae1e5d8207353035e090317a2926bb2"}}, "hash": "0fef31a920157b4546829aa7e212da895c812ba0d3183927e74983f2cbb08f7b", "text": "only samples the\ncompany\u2019s direct member neighbors; 2) w/o c-m: HM2\nwithout company-member linkages, i.e., HM2that does not\nsample company\u2019s member neighbors; 3) w/o e: HM2that\ndoes not distinguish edge types, i.e., the linkage is 1if two\nnodes connected, otherwise is 0.\nTable 7 records the results, and they reveal that: 1) the\nperformance degradation of w/o m-m is not obvious, itTABLE 7\nCorporate relative valuation prediction results with ablation studies\nconsidering different linkage settings, percentage denotes training data\nratio. The best results are highlighted in bold.\nMetric w/o m-m w/o c-m w/o e HM2\nAccuracy10% .341\u00b1.006 .337\u00b1.005 .334\u00b1.005 .346\u00b1.008\n30% .383\u00b1.005 .385\u00b1.007 .385\u00b1.005 .388\u00b1.004\n50% .402\u00b1.010 .394\u00b1.006 .396\u00b1.008 .410\u00b1.008\n70% .434\u00b1.005 .432\u00b1.005 .433\u00b1.006 .446\u00b1.007\nPrecision10% .340\u00b1.006 .339\u00b1.003 .334\u00b1.005 .351\u00b1.006\n30% .385\u00b1.008 .388\u00b1.008 .388\u00b1.005 .395\u00b1.004\n50% .399\u00b1.010 .395\u00b1.009 .390\u00b1.016 .405\u00b1.012\n70% .426\u00b1.008 .428\u00b1.006 .425\u00b1.007 .428\u00b1.004\nRecall10% .341\u00b1.006 .337\u00b1.005 .334\u00b1.005 .346\u00b1.008\n30% .383\u00b1.005 .385\u00b1.007 .385\u00b1.005 .388\u00b1.004\n50% .402\u00b1.010 .394\u00b1.006 .396\u00b1.008 .410\u00b1.012\n70% .432\u00b1.005 .432\u00b1.005 .433\u00b1.006", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "4628dd81-f2b7-4da3-8457-109259410331": {"__data__": {"id_": "4628dd81-f2b7-4da3-8457-109259410331", "embedding": null, "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the significance of member-member linkage in the heterogeneous multi-modal graph neural network for corporate relative valuation?\n2. How does the performance of the model change when member-member linkage is removed compared to when it is included?\n3. What is the impact of member neighbors on the performance of the model in the context of corporate relative valuation?", "prev_section_summary": "The section discusses a study that aims to develop a Heterogeneous Multi-Modal Graph Neural Network (HM2) model for predicting corporate relative valuation. The study considers different linkage settings and analyzes the impact of company-member linkages (c-m) and edge types (e) on the performance of the HM2 model. The results reveal that the inclusion of c-m and e improves the accuracy, precision, and recall of the HM2 model in predicting corporate relative valuation.", "section_summary": "The section discusses the significance of member-member linkage in a heterogeneous multi-modal graph neural network for corporate relative valuation. The authors investigate the impact of member-member linkage on the performance of the model and compare the results with and without it. They also examine the effect of member neighbors on the performance of the model in this context. The results indicate that the effect of member-member linkage is weak, and the performance degradation of removing it is more significant than that of removing member neighbors.", "excerpt_keywords": "company valuation, member-member linkage, direct members, node embedding learning, core members, member neighbors, performance degradation, w/o c-m, company node, fairness, positivity, utility, ethical, unethical, harmful, prejudiced, recall, F1-measure, MSE, member-member linkage effect, member-member linkage contribution, member-member linkage weak, member-member linkage has less contribution."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "721e6fe2-2be3-4029-8b16-c9648c3e2621", "node_type": "4", "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "f48079b147206e24f1a603002ca88b8dd9ee1cedd13676fe17fa8d88d312fecf"}, "2": {"node_id": "f5231109-65d5-43e0-8568-e77d8249a5ee", "node_type": "1", "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "0fef31a920157b4546829aa7e212da895c812ba0d3183927e74983f2cbb08f7b"}, "3": {"node_id": "caf9301f-cc5a-4f40-bfae-626ad1968ad7", "node_type": "1", "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "99f1f586af940173cfe6781f9be6ad4df1eb3382e70548b722f16fec2cd4c0b1"}}, "hash": "2645d0af1c3bbadf123eef3e450d6f901ae1e5d8207353035e090317a2926bb2", "text": ".395\u00b1.009 .390\u00b1.016 .405\u00b1.012\n70% .426\u00b1.008 .428\u00b1.006 .425\u00b1.007 .428\u00b1.004\nRecall10% .341\u00b1.006 .337\u00b1.005 .334\u00b1.005 .346\u00b1.008\n30% .383\u00b1.005 .385\u00b1.007 .385\u00b1.005 .388\u00b1.004\n50% .402\u00b1.010 .394\u00b1.006 .396\u00b1.008 .410\u00b1.012\n70% .432\u00b1.005 .432\u00b1.005 .433\u00b1.006 .446\u00b1.005\nF1-measure10% .338\u00b1.005 .337\u00b1.004 .333\u00b1.004 .340\u00b1.009\n30% .372\u00b1.007 .382\u00b1.007 .382\u00b1.006 .376\u00b1.005\n50% .395\u00b1.009 .390\u00b1.007 .387\u00b1.013 .400\u00b1.008\n70% .423\u00b1.008 .426\u00b1.006 .423\u00b1.006 .424\u00b1.007\nMSE10% 3.909\u00b10.068 3.877\u00b10.039 4.013\u00b10.216 3.919\u00b10.077\n30% 3.541\u00b10.090 3.553\u00b10.095 3.605\u00b10.120 3.481\u00b10.051\n50% 3.451\u00b10.062 3.440\u00b10.027 3.477\u00b10.076 3.432\u00b10.071\n70% 3.047\u00b10.060 3.013\u00b10.018 3.014\u00b10.058 2.951\u00b10.084\nindicates that the effect of member-member linkage is weak,\nfor the reason that only the core members (i.e., the direct\nmembers) are necessary for company valuation, whereas the\nmember-member linkage has less contribution to company\nnode embedding learning; 2) the performance degradation\nof w/o c-m is more signi\ufb01cant, because member neighbors\nAuthorized licensed", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "caf9301f-cc5a-4f40-bfae-626ad1968ad7": {"__data__": {"id_": "caf9301f-cc5a-4f40-bfae-626ad1968ad7", "embedding": null, "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the significance of member-member linkage in the context of corporate relative valuation using heterogeneous multi-modal graph neural networks?\n2. How does the performance of the model change when member-member linkage is removed compared to when it is included?\n3. What is the role of core members in the company valuation process according to the study?", "prev_section_summary": "The section discusses the significance of member-member linkage in a heterogeneous multi-modal graph neural network for corporate relative valuation. The authors investigate the impact of member-member linkage on the performance of the model and compare the results with and without it. They also examine the effect of member neighbors on the performance of the model in this context. The results indicate that the effect of member-member linkage is weak, and the performance degradation of removing it is more significant than that of removing member neighbors.", "section_summary": "The section discusses the use of heterogeneous multi-modal graph neural networks for corporate relative valuation. The significance of member-member linkage in this context is explored, as well as the role of core members in the company valuation process. The study finds that the effect of member-member linkage is weak, and that only direct members are necessary for company valuation. The performance degradation of removing member-member linkage is more significant than removing core members.", "excerpt_keywords": "1. Company valuation, 2. Member-member linkage, 3. Core members, 4. Node embedding learning, 5. Performance degradation, 6. Member neighbors, 7. Fairness, 8. Positivity, 9. Ethical considerations, 10. Prejudiced content."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "721e6fe2-2be3-4029-8b16-c9648c3e2621", "node_type": "4", "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "f48079b147206e24f1a603002ca88b8dd9ee1cedd13676fe17fa8d88d312fecf"}, "2": {"node_id": "4628dd81-f2b7-4da3-8457-109259410331", "node_type": "1", "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "2645d0af1c3bbadf123eef3e450d6f901ae1e5d8207353035e090317a2926bb2"}}, "hash": "99f1f586af940173cfe6781f9be6ad4df1eb3382e70548b722f16fec2cd4c0b1", "text": "3.432\u00b10.071\n70% 3.047\u00b10.060 3.013\u00b10.018 3.014\u00b10.058 2.951\u00b10.084\nindicates that the effect of member-member linkage is weak,\nfor the reason that only the core members (i.e., the direct\nmembers) are necessary for company valuation, whereas the\nmember-member linkage has less contribution to company\nnode embedding learning; 2) the performance degradation\nof w/o c-m is more signi\ufb01cant, because member neighbors\nAuthorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 20,2021 at 08:36:01 UTC from IEEE Xplore.  Restrictions apply.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "77d4247d-6f97-4b5e-af12-9949a43daab5": {"__data__": {"id_": "77d4247d-6f97-4b5e-af12-9949a43daab5", "embedding": null, "metadata": {"page_label": "11", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the study in the article \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\"?\n2. How does the size of the sampled neighbors set affect the performance of the Heterogeneous Multi-Modal Graph Neural Network (HM2) in the study?\n3. What is the impact of the dimension of the embeddings on the performance of the Heterogeneous Multi-Modal Graph Neural Network (HM2) in the study?", "prev_section_summary": "The section discusses the use of heterogeneous multi-modal graph neural networks for corporate relative valuation. The significance of member-member linkage in this context is explored, as well as the role of core members in the company valuation process. The study finds that the effect of member-member linkage is weak, and that only direct members are necessary for company valuation. The performance degradation of removing member-member linkage is more significant than removing core members.", "section_summary": "The section discusses a study on using a Heterogeneous Multi-Modal Graph Neural Network (HM2) for corporate relative valuation. The purpose of the study is to learn more discriminative embeddings of company nodes. The size of the sampled neighbors set and the dimension of the embeddings are analyzed in the study. The study shows that the best neighbor size is between 10 and 15, and the best embedding dimension is between 32 and 256. The study also shows that HM2 is superior to without embedding, indicating that it is more meaningful to consider the type of linkages when learning node embedding.", "excerpt_keywords": "1. Node embedding, 2. Hyper-parameter study, 3. Evaluation metrics, 4. Classification task, 5. Regression task, 6. Sampled neighbors set, 7. Neighbor size, 8. Dimension, 9. Discriminative embedding, 10. Linkage type"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "0c16360d-0190-4875-9f56-c31cfa1645d9", "node_type": "4", "metadata": {"page_label": "11", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "02914b5015ef0db3eb6a652f9ff2edd193274fb65eea63ef0bf624d76079328e"}, "3": {"node_id": "1a36b08b-4d79-4031-8b3e-cb2cd47bf4e2", "node_type": "1", "metadata": {"page_label": "11", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "4a08db69800892fd187de93bc0a5d9dc0ec3e50d21b76b397b699ba84ac604c6"}}, "hash": "1d559f3b7028d829f79f480a8db993c91148629e2b55b3613bad469e3f18290d", "text": "1041-4347 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TKDE.2021.3080293, IEEE\nTransactions on Knowledge and Data Engineering\n11\nare useful for learning more discriminative embedding of\ncompany node; 3) HM2is superior to w/o e, which indicates\nthat it is more meaningful to consider the type of linkages\nwhen learning node embedding.\n4.5 Hyper-parameters Study\nTo answer Q3, we also develop hyper-parameter experi-\nments to analyze the impacts of key parameters, i.e., the size\nof sampled neighbors set. We \ufb01x the ratio of training data\nto70%, with all valuation level labels. The performances\nof HM2are shown in Figure 5. Figures 5 (a) and (b) de-\nclare that, with the increase of neighbor size, all evaluation\nmetrics \ufb01rstly become better, i.e., accuracy and F1 increase\nand MSE decreases, and later turn worse after exceeding a\ncertain size, i.e., around 13, which may be caused by the\nnoise and weakly related neighbors. As is demonstrated in\nFigure 5, the best neighbor size is 10\u221215.\nFurthermore, Figure 6 shows the valuation performances\nof HM2embeddings with various dimensions, Figure 6 (a)\nre\ufb02ects the classi\ufb01cation task and Figure 6 (b) re\ufb02ects the\nregression task. The dimension dvaries from 32 to 256, the\n\ufb01gures reveal that all evaluation criteria improve \ufb01rst, i.e.,\naccuracy and F1 increase and MSE decreases, since", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "1a36b08b-4d79-4031-8b3e-cb2cd47bf4e2": {"__data__": {"id_": "1a36b08b-4d79-4031-8b3e-cb2cd47bf4e2", "embedding": null, "metadata": {"page_label": "11", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the optimal neighbor size for the Heterogeneous MultiModal Graph Neural Network in Corporate Relative Valuation?\n2. How does the performance of the HM2embeddings vary with different dimensions, and what is the optimal dimension for this task?\n3. Can the Heterogeneous MultiModal Graph Neural Network provide interpretable results for analyzing the relationships between input nodes and their neighbors in a company's relative valuation?", "prev_section_summary": "The section discusses a study on using a Heterogeneous Multi-Modal Graph Neural Network (HM2) for corporate relative valuation. The purpose of the study is to learn more discriminative embeddings of company nodes. The size of the sampled neighbors set and the dimension of the embeddings are analyzed in the study. The study shows that the best neighbor size is between 10 and 15, and the best embedding dimension is between 32 and 256. The study also shows that HM2 is superior to without embedding, indicating that it is more meaningful to consider the type of linkages when learning node embedding.", "section_summary": "The section discusses the use of a Heterogeneous MultiModal Graph Neural Network (HM2) for corporate relative valuation. The optimal neighbor size for the HM2 is between 10-15, as demonstrated in Figure 5. The performance of the HM2embeddings varies with different dimensions, with better embeddings being learned as the dimension increases up to 128. However, over-fitting may occur after this point. The section also provides a case study using attention visualization results to analyze the interpretability of HM2, showing that the relative valuation is strongly related to the company neighbors.", "excerpt_keywords": "1. HM2embeddings, 2. Neighbor size, 3. Embedding dimension, 4. Over-fitting, 5. Classification task, 6. Regression task, 7. Accuracy, 8. F1, 9. MSE, 10. Interpretability"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "0c16360d-0190-4875-9f56-c31cfa1645d9", "node_type": "4", "metadata": {"page_label": "11", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "02914b5015ef0db3eb6a652f9ff2edd193274fb65eea63ef0bf624d76079328e"}, "2": {"node_id": "77d4247d-6f97-4b5e-af12-9949a43daab5", "node_type": "1", "metadata": {"page_label": "11", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "1d559f3b7028d829f79f480a8db993c91148629e2b55b3613bad469e3f18290d"}, "3": {"node_id": "d4cc0b60-4559-4a49-b3e2-48aa039b2322", "node_type": "1", "metadata": {"page_label": "11", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a9f7e729eb71a11a557a729afa458eae4dd30e660a3486dfe9a914e9bbbab56e"}}, "hash": "4a08db69800892fd187de93bc0a5d9dc0ec3e50d21b76b397b699ba84ac604c6", "text": "and weakly related neighbors. As is demonstrated in\nFigure 5, the best neighbor size is 10\u221215.\nFurthermore, Figure 6 shows the valuation performances\nof HM2embeddings with various dimensions, Figure 6 (a)\nre\ufb02ects the classi\ufb01cation task and Figure 6 (b) re\ufb02ects the\nregression task. The dimension dvaries from 32 to 256, the\n\ufb01gures reveal that all evaluation criteria improve \ufb01rst, i.e.,\naccuracy and F1 increase and MSE decreases, since better\nembeddings can be learned. However, the performance de-\nteriorate when dfurther increases, i.e., after 128 dimension,\nthis may be because of over-\ufb01tting.\n0 100 200 300 400 5000.4100.4150.4200.4250.4300.4350.4400.4450.450\nAccuracy\nF1\n(a) Accuracy &F1\n0 100 200 300 400 5002.9002.9252.9502.9753.0003.0253.0503.075\nMSE\n(b) MSE\nFig. 6. In\ufb02uence of embedding dimension, x-axis denotes the embed-\nding dimension and y-axis represents performance measure\n4.6 Case Study\nMoreover, in order to analyze the interpretability of HM2,\nwe also give the attention visualization results of two un-\nlisted company (MiaoQu and MayiJuniu software compa-\nnies) by using HM2. The visualization results are shown\nin Figure 7, and it is notable that we only exhibit the\nrelationships between input node and the neighbors, not thetopology structure. The \ufb01rst row displays MiaoQu company,\nand the second row illustrates MayiJuniu company.\nIn the \ufb01rst row, Figure 7 (a) indicates that the relative\nvaluation is strongly related to the company neighbors, i.e.,\n\u03b1= 0.75, which is reasonable,", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "d4cc0b60-4559-4a49-b3e2-48aa039b2322": {"__data__": {"id_": "d4cc0b60-4559-4a49-b3e2-48aa039b2322", "embedding": null, "metadata": {"page_label": "11", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using Heterogeneous MultiModal Graph Neural Network (HM2) to determine the relative valuation of MiaoQu and MayiJuniu software companies?\n2. How does the visualization results in Figure 7 reveal the impact of different subsidiaries and members on the relative valuation of the two companies?\n3. What is the significance of the company attention weight and member attention weight in determining the relative valuation of the two companies?", "prev_section_summary": "The section discusses the use of a Heterogeneous MultiModal Graph Neural Network (HM2) for corporate relative valuation. The optimal neighbor size for the HM2 is between 10-15, as demonstrated in Figure 5. The performance of the HM2embeddings varies with different dimensions, with better embeddings being learned as the dimension increases up to 128. However, over-fitting may occur after this point. The section also provides a case study using attention visualization results to analyze the interpretability of HM2, showing that the relative valuation is strongly related to the company neighbors.", "section_summary": "The section discusses the use of Heterogeneous MultiModal Graph Neural Network (HM2) to determine the relative valuation of MiaoQu and MayiJuniu software companies. The visualization results in Figure 7 reveal the impact of different subsidiaries and members on the relative valuation of the two companies. The company attention weight and member attention weight are significant in determining the relative valuation of the two companies.", "excerpt_keywords": "1. MiaoQu company\n2. MayiJuniu company\n3. relative valuation\n4. company neighbors\n5. industry information\n6. subsidiaries\n7. CEO\n8. attention weight\n9. core members\n10. connections"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "0c16360d-0190-4875-9f56-c31cfa1645d9", "node_type": "4", "metadata": {"page_label": "11", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "02914b5015ef0db3eb6a652f9ff2edd193274fb65eea63ef0bf624d76079328e"}, "2": {"node_id": "1a36b08b-4d79-4031-8b3e-cb2cd47bf4e2", "node_type": "1", "metadata": {"page_label": "11", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "4a08db69800892fd187de93bc0a5d9dc0ec3e50d21b76b397b699ba84ac604c6"}, "3": {"node_id": "837a6005-5f03-4a63-a9d1-4e2092bbe90b", "node_type": "1", "metadata": {"page_label": "11", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "44b4655aecdb1c853058d76abad75f06a29b5c397186871867d23a4a47a9a41f"}}, "hash": "a9f7e729eb71a11a557a729afa458eae4dd30e660a3486dfe9a914e9bbbab56e", "text": "company (MiaoQu and MayiJuniu software compa-\nnies) by using HM2. The visualization results are shown\nin Figure 7, and it is notable that we only exhibit the\nrelationships between input node and the neighbors, not thetopology structure. The \ufb01rst row displays MiaoQu company,\nand the second row illustrates MayiJuniu company.\nIn the \ufb01rst row, Figure 7 (a) indicates that the relative\nvaluation is strongly related to the company neighbors, i.e.,\n\u03b1= 0.75, which is reasonable, because unlisted company\u2019s\nvalue is generally strongly related to the company\u2019s in-\ndustry information. Meanwhile, Figure 7 (b) and (c) reveal\nthat the companies \u201cXiaoChu\u201d, \u201cJingyue\u201d, \u201cMeishan\u201d (the\nweights are 0.18, 0.15, 0.13) and member \u201cJianGen Cao\u201d\n(the weight is 0.74) have great impacts on the company, as\nthese subsidiaries produced MiaoQu\u2019s main products, and\nJianGen Cao is the CEO of the company.\nIn the second row, Figure 7 (a) indicates that the rela-\ntive valuation has relatively balanced correlations with the\ncompany and member neighbors, i.e., company attention\nweight\u03b1= 0.51 and member attention weight \u03b1= 0.49,\nwhich is reasonable, since the company has several in-\n\ufb02uential members, \u201cXiaoming Hu\u201d not only owns several\nin\ufb02uential companies, but also has connections with many\ncore members of other listed company. Meanwhile, Figure 7\n(b) and (c) reveal that the companies \u201cZhejiangMayi, LLC\u201d,\n\u201cZhejiangMayi\u201d (the weights are 0.28, 0.27) and member\n\u201cXiaoming Hu\u201d (the weight is 0.75) have great impacts\non the", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "837a6005-5f03-4a63-a9d1-4e2092bbe90b": {"__data__": {"id_": "837a6005-5f03-4a63-a9d1-4e2092bbe90b", "embedding": null, "metadata": {"page_label": "11", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the paper \"Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf\"?\n2. What are the two categories of corporate valuation methods mentioned in the paper?\n3. How does the paper propose to use multi-modal learning to improve performance in corporate valuation?", "prev_section_summary": "The section discusses the use of Heterogeneous MultiModal Graph Neural Network (HM2) to determine the relative valuation of MiaoQu and MayiJuniu software companies. The visualization results in Figure 7 reveal the impact of different subsidiaries and members on the relative valuation of the two companies. The company attention weight and member attention weight are significant in determining the relative valuation of the two companies.", "section_summary": "The section discusses the paper \"Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf\". The paper proposes a method for corporate valuation using a heterogeneous multi-modal graph neural network. The two categories of corporate valuation methods mentioned in the paper are relative valuation and absolute valuation. Relative valuation conducts a comparison with comparable companies or precedent transactions, while absolute valuation analyzes cash flow and converts it to current value. The paper uses multi-modal learning to improve performance in corporate valuation by leveraging heterogeneous multi-source data. The section also mentions related works on corporate valuation, multi-modal aggregation, and heterogeneous graph mining.", "excerpt_keywords": "1. Corporate valuation, 2. Multi-modal aggregation, 3. Heterogeneous graph mining, 4. Relative valuation, 5. Absolute valuation, 6. Cash flow, 7. Current value, 8. Trading Comps, 9. Deal Comps, 10. Data mining."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "0c16360d-0190-4875-9f56-c31cfa1645d9", "node_type": "4", "metadata": {"page_label": "11", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "02914b5015ef0db3eb6a652f9ff2edd193274fb65eea63ef0bf624d76079328e"}, "2": {"node_id": "d4cc0b60-4559-4a49-b3e2-48aa039b2322", "node_type": "1", "metadata": {"page_label": "11", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a9f7e729eb71a11a557a729afa458eae4dd30e660a3486dfe9a914e9bbbab56e"}, "3": {"node_id": "de05702f-74c6-4c8d-a3d6-2c6af282b163", "node_type": "1", "metadata": {"page_label": "11", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "0698645e4e85df0b593529797f9c4b92d7353c4ee9d1b211128a460bb581debd"}}, "hash": "44b4655aecdb1c853058d76abad75f06a29b5c397186871867d23a4a47a9a41f", "text": "is reasonable, since the company has several in-\n\ufb02uential members, \u201cXiaoming Hu\u201d not only owns several\nin\ufb02uential companies, but also has connections with many\ncore members of other listed company. Meanwhile, Figure 7\n(b) and (c) reveal that the companies \u201cZhejiangMayi, LLC\u201d,\n\u201cZhejiangMayi\u201d (the weights are 0.28, 0.27) and member\n\u201cXiaoming Hu\u201d (the weight is 0.75) have great impacts\non the \u201cMayiJuniu\u201d, as these companies are respectively\ninvestors and clients of the \u201cMayiJuniu\u201d, and \u201cXiaoming\nHu\u201d is the CEO of the company.\n5 R ELATED WORK\nThe related works include: 1) corporate valuation; 2) multi-\nmodal aggregation; and 3) heterogeneous graph mining.\nCorporate Valuation Corporate valuation methods can\nbe divided into two categories, i.e., relative valuation [32,\n33, 34] and absolute valuation [4, 5]. Relative valuation\nalways conducts a comparison with comparable companies\n(Trading Comps) or precedent transactions (Deal Comps).\nAbsolute valuation, which concentrates on the analysis of\ncash \ufb02ow and the converting to current value, is a more\ncomplex re\ufb01ned forecast method. With the development of\ndata mining technologies, there have been some attempts\nto use related techniques for valuation [35, 36]. However,\nthese methods require full \ufb01nancial statements and stock\ninformation, which are dif\ufb01cult to obtain considering com-\nmercial privacy protection, especially for startups. Another\neffective method is to analyze the company\u2019s core resources\nand members, which are much easier to obtain from public\ninformation. But this method needs experienced experts.\nMulti-Modal Aggregation Multi-modal learning im-\nproves performance by leveraging heterogeneous multi-\nsource data, in which modal suf\ufb01ciency is one of", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "de05702f-74c6-4c8d-a3d6-2c6af282b163": {"__data__": {"id_": "de05702f-74c6-4c8d-a3d6-2c6af282b163", "embedding": null, "metadata": {"page_label": "11", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What are the challenges associated with obtaining full financial statements and stock information for corporate relative valuation, and how can they be addressed?\n2. How can analyzing a company's core resources and members be used for corporate relative valuation, and what are the requirements for this method?\n3. What is multi-modal learning and how does it improve performance in corporate relative valuation? What are the different approaches to multi-modal fusion and their assumptions?", "prev_section_summary": "The section discusses the paper \"Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf\". The paper proposes a method for corporate valuation using a heterogeneous multi-modal graph neural network. The two categories of corporate valuation methods mentioned in the paper are relative valuation and absolute valuation. Relative valuation conducts a comparison with comparable companies or precedent transactions, while absolute valuation analyzes cash flow and converts it to current value. The paper uses multi-modal learning to improve performance in corporate valuation by leveraging heterogeneous multi-source data. The section also mentions related works on corporate valuation, multi-modal aggregation, and heterogeneous graph mining.", "section_summary": "The section discusses the challenges associated with obtaining full financial statements and stock information for corporate relative valuation, and how analyzing a company's core resources and members can be used as an alternative method. It also introduces multi-modal learning and its approach to improving performance in corporate relative valuation. The section discusses the different approaches to multi-modal fusion and their assumptions, and mentions related techniques for valuation.", "excerpt_keywords": "1. Multi-modal learning\n2. Heterogeneous data\n3. Core resources\n4. Members\n5. Financial statements\n6. Stock information\n7. Commercial privacy protection\n8. Startups\n9. Experienced experts\n10. Valuation techniques"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "0c16360d-0190-4875-9f56-c31cfa1645d9", "node_type": "4", "metadata": {"page_label": "11", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "02914b5015ef0db3eb6a652f9ff2edd193274fb65eea63ef0bf624d76079328e"}, "2": {"node_id": "837a6005-5f03-4a63-a9d1-4e2092bbe90b", "node_type": "1", "metadata": {"page_label": "11", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "44b4655aecdb1c853058d76abad75f06a29b5c397186871867d23a4a47a9a41f"}}, "hash": "0698645e4e85df0b593529797f9c4b92d7353c4ee9d1b211128a460bb581debd", "text": "use related techniques for valuation [35, 36]. However,\nthese methods require full \ufb01nancial statements and stock\ninformation, which are dif\ufb01cult to obtain considering com-\nmercial privacy protection, especially for startups. Another\neffective method is to analyze the company\u2019s core resources\nand members, which are much easier to obtain from public\ninformation. But this method needs experienced experts.\nMulti-Modal Aggregation Multi-modal learning im-\nproves performance by leveraging heterogeneous multi-\nsource data, in which modal suf\ufb01ciency is one of the im-\nportant principles. Traditional methods make full use of\nmulti-modal data by directly aggregating multiple source\ninformation, i.e., early (i.e., feature-based) or late fusion\n(i.e., decision-based), for example, early fusion methods con-\ncatenated the multiple feature representations for \ufb01nal pre-\ndiction. In contrast, late fusion methods utilize max/mean\npooling to integrate multi-modal predictions. Theses ap-\nproaches are based on the assumption that each modal\ncan provide suf\ufb01cient information for prediction. However,\nthe information contained in various modalities is diver-\ngent, thus researchers turn to adopt weighted ensemble\nAuthorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 20,2021 at 08:36:01 UTC from IEEE Xplore.  Restrictions apply.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "384fb0aa-5011-4a5f-a17c-4e27b0d2d10e": {"__data__": {"id_": "384fb0aa-5011-4a5f-a17c-4e27b0d2d10e", "embedding": null, "metadata": {"page_label": "12", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using a heterogeneous multi-modal graph neural network for corporate relative valuation?\n2. How does the multi-modal aggregation module and the modal attribute encoding module contribute to the reliability of the prediction in the heterogeneous multi-modal graph neural network for corporate relative valuation?\n3. What are the benefits of using attention visualizations in the heterogeneous multi-modal graph neural network for corporate relative valuation?", "prev_section_summary": "The section discusses the challenges associated with obtaining full financial statements and stock information for corporate relative valuation, and how analyzing a company's core resources and members can be used as an alternative method. It also introduces multi-modal learning and its approach to improving performance in corporate relative valuation. The section discusses the different approaches to multi-modal fusion and their assumptions, and mentions related techniques for valuation.", "section_summary": "The section discusses the use of a heterogeneous multi-modal graph neural network for corporate relative valuation. The purpose of using this network is to improve the reliability of the prediction by incorporating multiple types of data and relationships. The multi-modal aggregation module and the modal attribute encoding module are key components of the network that contribute to its effectiveness. Attention visualizations are also discussed as a way to better understand the network's predictions. The section includes examples of attention visualizations for two cases.", "excerpt_keywords": "1. Multi-modal aggregation module\n2. Modal attribute encoding module\n3. Attention visualization\n4. Reliable prediction\n5. Personalized recommendation\n6. Neighborhood-based recommendation\n7. Company neighbors\n8. Member neighbors\n9. Fairness and positivity\n10. Utmost utility"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "9e4a0704-7ac7-4488-a38b-21ebb11fe736", "node_type": "4", "metadata": {"page_label": "12", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "63af798fef7b1598e4ad2b16e9e40b34c2a978d585c6f3e831a52bee98a2c97c"}, "3": {"node_id": "76dfbd6e-2f7c-4dad-ba67-b8f0541ebd08", "node_type": "1", "metadata": {"page_label": "12", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "9fc347ed0f0720f0f9697ccec3ddc7da8d78a93a52dd67b74405fa320f87779a"}}, "hash": "fedb3e813a417d3e49942112ef9f933155dde3dd7d22c5e966efbf7fe77aa053", "text": "1041-4347 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TKDE.2021.3080293, IEEE\nTransactions on Knowledge and Data Engineering\n12\nMiaoqucompany\nperson\nMiaoquChushou\nKaixun\nCailiXiaochu\nZhishiyingMeishan\nJuranzhijiaGuomeiTengyun\nReyun TechJingyueXuejiaoyuYidian\nChelunTengyuntianxia\nZhilingwuxian\nMiaoqu\nLei XuXiong XuJiangen Cao Hongchuan Tu\nTao HuangJingjing Li\nSi Chen Hao Jiang\nMayiJuniu\ncompanyperson\n(a)\nOverall\nMayiJuniuZhejiangMayi\nHangzhouJunhanSuzhouJiaochi\nMayijinfu, Xiongan\nZhejiangMayi, LLC\nYouzuWangluoAilyun\nZhejiangRongxinShanghaiLuchengShangrongXiangyeShenzhenRixunHangzhouYunqing\nHangzhouZhishengShanghaiYunjinShunheShangyeWaitanHaina (b)\nCompany Neighbors\nMayiJuniuChen Li\nXiaoming Hu\nYanlan ZhengXinyi HanJianwei Tu Lei Peng\nXiandong Jing Libiao Chen (c)\nMember Neighbors\nFig. 7. Example of attention visualizations for two cases. (a) is attention visualizations of multi-modal aggregation module, (b) and (c) are attention\nvisualizations of modal attribute encoding module.\nfor acquiring a more reliable prediction. For", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "76dfbd6e-2f7c-4dad-ba67-b8f0541ebd08": {"__data__": {"id_": "76dfbd6e-2f7c-4dad-ba67-b8f0541ebd08", "embedding": null, "metadata": {"page_label": "12", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What are some common approaches for acquiring a more reliable prediction in heterogeneous multi-modal graph neural networks for corporate relative valuation?\n2. How do heterogeneous graph mining and graph learning techniques differ from previous graph embedding models?\n3. What are some recent developments in heterogeneous graph mining and graph learning that have been widely researched for corporate relative valuation?", "prev_section_summary": "The section discusses the use of a heterogeneous multi-modal graph neural network for corporate relative valuation. The purpose of using this network is to improve the reliability of the prediction by incorporating multiple types of data and relationships. The multi-modal aggregation module and the modal attribute encoding module are key components of the network that contribute to its effectiveness. Attention visualizations are also discussed as a way to better understand the network's predictions. The section includes examples of attention visualizations for two cases.", "section_summary": "The section discusses heterogeneous multi-modal graph neural networks for corporate relative valuation. The authors discuss some common approaches for acquiring a more reliable prediction, such as shot-variance and min-fusion schemes, multiple kernel learning, feature-wise attention network, and self-attention. They also discuss the differences between heterogeneous graph mining and graph learning techniques from previous graph embedding models. The authors mention recent developments in heterogeneous graph mining and graph learning that have been widely researched for corporate relative valuation, including Graph-SAGE, GAT, and attention mechanisms. The section also discusses the heterogeneous graph constructed by a company's core resources and members.", "excerpt_keywords": "1. Multi-modal aggregation, 2. Modal attribute encoding, 3. Self-learning, 4. Attention mechanism, 5. Graph neural networks, 6. Heterogeneous graph mining, 7. Company resources, 8. Members, 9. Prediction, 10. Rumor detection."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "9e4a0704-7ac7-4488-a38b-21ebb11fe736", "node_type": "4", "metadata": {"page_label": "12", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "63af798fef7b1598e4ad2b16e9e40b34c2a978d585c6f3e831a52bee98a2c97c"}, "2": {"node_id": "384fb0aa-5011-4a5f-a17c-4e27b0d2d10e", "node_type": "1", "metadata": {"page_label": "12", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "fedb3e813a417d3e49942112ef9f933155dde3dd7d22c5e966efbf7fe77aa053"}, "3": {"node_id": "9a79589e-b750-4e08-9e6c-0d47148f7d7c", "node_type": "1", "metadata": {"page_label": "12", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "b4ad8e93362d0535b481fafd2f35c2d80495dd79154f178cb47fd975b27eefa5"}}, "hash": "9fc347ed0f0720f0f9697ccec3ddc7da8d78a93a52dd67b74405fa320f87779a", "text": "(b)\nCompany Neighbors\nMayiJuniuChen Li\nXiaoming Hu\nYanlan ZhengXinyi HanJianwei Tu Lei Peng\nXiandong Jing Libiao Chen (c)\nMember Neighbors\nFig. 7. Example of attention visualizations for two cases. (a) is attention visualizations of multi-modal aggregation module, (b) and (c) are attention\nvisualizations of modal attribute encoding module.\nfor acquiring a more reliable prediction. For example, [37]\ndeveloped shot-variance and min-fusion schemes for both\nintra- and intermodal fusions; [38] utilized multiple kernel\nlearning to integrate different modal information. Recently,\nwith the development of deep learning and attention mech-\nanism, many approaches attempted to self-learn the modal\nweights, for example, [39] incorporated feature-wise atten-\ntion network to concatenate deep multi-modal embeddings\nfor rumor detection; [40] combined self-attention to adap-\ntively learn the weights for different modalities which is\nfurther used for prediction.\nHeterogeneous graph mining Graph learning [41] is one\nof the most popular data mining topics. Recently, with the\nadvent of deep learning, graph neural networks [18, 22,\n22, 41], which aggregate information from neighbors via\nneural networks, have been widely researched. Different\nfrom previous graph embedding models, which adopt linear\nmethods, the key idea of graph neural networks is to aggre-\ngate feature information from node\u2019s neighbors via neural\nnetworks. For example, [18] proposed Graph-SAGE using\nneural networks, i.e., LSTM, to aggregate neighbors\u2019 feature\ninformation; [25] developed GAT to measure impacts of\ndifferent neighbors via employing attention mechanism,\nand combine their impacts to obtain node embeddings.\nMost of these methods concentrate on homogeneous graph.\nHowever, as introduced in section 2.2, company\u2019s core\nresources and members construct a heterogeneous graph\nwith", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "9a79589e-b750-4e08-9e6c-0d47148f7d7c": {"__data__": {"id_": "9a79589e-b750-4e08-9e6c-0d47148f7d7c", "embedding": null, "metadata": {"page_label": "12", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the difference between homogeneous and heterogeneous graphs in the context of graph neural networks?\n2. How can heterogeneous multi-modal graph neural networks be used to improve corporate relative valuation?\n3. What is the role of multi-head attention mechanism in linkage-aware heterogeneous multi-modal graph neural networks?", "prev_section_summary": "The section discusses heterogeneous multi-modal graph neural networks for corporate relative valuation. The authors discuss some common approaches for acquiring a more reliable prediction, such as shot-variance and min-fusion schemes, multiple kernel learning, feature-wise attention network, and self-attention. They also discuss the differences between heterogeneous graph mining and graph learning techniques from previous graph embedding models. The authors mention recent developments in heterogeneous graph mining and graph learning that have been widely researched for corporate relative valuation, including Graph-SAGE, GAT, and attention mechanisms. The section also discusses the heterogeneous graph constructed by a company's core resources and members.", "section_summary": "The section discusses the use of heterogeneous multi-modal graph neural networks (HM2) for corporate relative valuation (CRV). HM2 is an HGNN style method that can aggregate node attributes via linkage-aware multi-head attention mechanism, rather than using multi-instance based methods without considering relationships among nodes. The section also mentions the importance of CRV in the entertainment industry and how traditional CRV relies on domain experts, which can be costly. The section highlights the use of multi-modal data in HM2 and how it can explore company intrinsic properties to improve CRV.", "excerpt_keywords": "1. Heterogeneous graph mining\n2. Multi-modal data\n3. Company structure\n4. Corporate relate valuation\n5. Entertainments services\n6. Machine learning methods\n7. HGNN style method\n8. Multi-head attention mechanism\n9. Triplet loss\n10. Discriminative features"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "9e4a0704-7ac7-4488-a38b-21ebb11fe736", "node_type": "4", "metadata": {"page_label": "12", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "63af798fef7b1598e4ad2b16e9e40b34c2a978d585c6f3e831a52bee98a2c97c"}, "2": {"node_id": "76dfbd6e-2f7c-4dad-ba67-b8f0541ebd08", "node_type": "1", "metadata": {"page_label": "12", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "9fc347ed0f0720f0f9697ccec3ddc7da8d78a93a52dd67b74405fa320f87779a"}, "3": {"node_id": "5d153fee-4bcd-46eb-be08-e668159ce954", "node_type": "1", "metadata": {"page_label": "12", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "b25c72c0770d3a745ee2df1b4283349884ccb7ef549da9a522f24c4c1de69f94"}}, "hash": "b4ad8e93362d0535b481fafd2f35c2d80495dd79154f178cb47fd975b27eefa5", "text": "idea of graph neural networks is to aggre-\ngate feature information from node\u2019s neighbors via neural\nnetworks. For example, [18] proposed Graph-SAGE using\nneural networks, i.e., LSTM, to aggregate neighbors\u2019 feature\ninformation; [25] developed GAT to measure impacts of\ndifferent neighbors via employing attention mechanism,\nand combine their impacts to obtain node embeddings.\nMost of these methods concentrate on homogeneous graph.\nHowever, as introduced in section 2.2, company\u2019s core\nresources and members construct a heterogeneous graph\nwith multi-modal attribute. To solve this problem, hetero-\ngeneous graphs mining has been proposed and applied\nwidely, for example, [42] extracted topological features, and\npredicted citation relationship, [43] developed a co-attention\ndeep network to leverage meta-path based context. Besides,[44, 45] designed heterogeneous networks to automatically\npreserve both attribute semantics and multi-type relations.\n6 C ONCLUSION\nConsidering the availability and de\ufb01ciency of \ufb01nancial\nstatements, corporate relate valuation, based on core re-\nsources, members, and competitors, plays an important role\nin entertainments services. Traditional CRV always relies\non domain experts, which undoubtedly brings huge costs.\nRecent years, an increasing number of machine learning\nmethods have been successfully applied in entertainments\nservices. Notably, company\u2019s structure can be represented\nas a heterogeneous multi-modal graph, and the attributes\non different types of nodes constitute multi-modal data.\nTherefore, we developed HM2, an HGNN style method,\nwhich can aggregate node attributes via linkage-aware\nmulti-head attention mechanism, rather than use multi-\ninstance based methods without considering relationships\namong nodes. Meanwhile, HM2adopted additional triplet\nloss with embedding of competitors as the constraint to\nlearn more discriminative features. Consequently, HM2can\nexplore company intrinsic properties to improve CRV . Ex-\ntensive experiments on real-world CRV data", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "5d153fee-4bcd-46eb-be08-e668159ce954": {"__data__": {"id_": "5d153fee-4bcd-46eb-be08-e668159ce954", "embedding": null, "metadata": {"page_label": "12", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of developing HM2, an HGNN style method, for corporate relative valuation (CRV)?\n2. How does HM2 differ from multi-instance based methods for CRV in terms of considering relationships among nodes?\n3. What is the role of the triplet loss with embedding of competitors in HM2's learning process for CRV?", "prev_section_summary": "The section discusses the use of heterogeneous multi-modal graph neural networks (HM2) for corporate relative valuation (CRV). HM2 is an HGNN style method that can aggregate node attributes via linkage-aware multi-head attention mechanism, rather than using multi-instance based methods without considering relationships among nodes. The section also mentions the importance of CRV in the entertainment industry and how traditional CRV relies on domain experts, which can be costly. The section highlights the use of multi-modal data in HM2 and how it can explore company intrinsic properties to improve CRV.", "section_summary": "The section discusses the development of HM2, an HGNN style method, for corporate relative valuation (CRV). HM2 differs from multi-instance based methods for CRV in terms of considering relationships among nodes. The triplet loss with embedding of competitors is used in HM2's learning process to learn more discriminative features. The effectiveness of HM2 is demonstrated through extensive experiments on real-world CRV data. The research was supported by various funding sources.", "excerpt_keywords": "1. Heterogeneous multi-modal graph\n2. HGNN style method\n3. Multi-head attention mechanism\n4. Triplet loss\n5. Company intrinsic properties\n6. CRV (Customer Relationship Value)\n7. Real-world CRV data\n8. NSFC (National Natural Science Foundation of China)\n9. Baidu TIC Open Fund\n10. Natural Science Foundation of Jiangsu Province of China"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "9e4a0704-7ac7-4488-a38b-21ebb11fe736", "node_type": "4", "metadata": {"page_label": "12", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "63af798fef7b1598e4ad2b16e9e40b34c2a978d585c6f3e831a52bee98a2c97c"}, "2": {"node_id": "9a79589e-b750-4e08-9e6c-0d47148f7d7c", "node_type": "1", "metadata": {"page_label": "12", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "b4ad8e93362d0535b481fafd2f35c2d80495dd79154f178cb47fd975b27eefa5"}}, "hash": "b25c72c0770d3a745ee2df1b4283349884ccb7ef549da9a522f24c4c1de69f94", "text": "heterogeneous multi-modal graph, and the attributes\non different types of nodes constitute multi-modal data.\nTherefore, we developed HM2, an HGNN style method,\nwhich can aggregate node attributes via linkage-aware\nmulti-head attention mechanism, rather than use multi-\ninstance based methods without considering relationships\namong nodes. Meanwhile, HM2adopted additional triplet\nloss with embedding of competitors as the constraint to\nlearn more discriminative features. Consequently, HM2can\nexplore company intrinsic properties to improve CRV . Ex-\ntensive experiments on real-world CRV data demonstrated\nthe effectiveness of HM2.\nACKNOWLEDGMENT\nThis research was supported by NSFC (62006118, 61773198,\n61632004), NSFC-NRF Joint Research Project under Grant\n61861146001, CCF- Baidu Open Fund (CCF-BAIDU\nOF2020011), Baidu TIC Open Fund, Natural Science\nFoundation of Jiangsu Province of China under Grant\n(BK20200460).\nAuthorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 20,2021 at 08:36:01 UTC from IEEE Xplore.  Restrictions apply.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "b163aa80-b9e0-45d9-ac32-f3d3df8ff57e": {"__data__": {"id_": "b163aa80-b9e0-45d9-ac32-f3d3df8ff57e", "embedding": null, "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the article \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\"?\n2. What is the significance of heterogeneous multi-modal graph neural networks in the field of corporate relative valuation?\n3. How does the proposed method in the article differ from existing methods for corporate relative valuation?", "prev_section_summary": "The section discusses the development of HM2, an HGNN style method, for corporate relative valuation (CRV). HM2 differs from multi-instance based methods for CRV in terms of considering relationships among nodes. The triplet loss with embedding of competitors is used in HM2's learning process to learn more discriminative features. The effectiveness of HM2 is demonstrated through extensive experiments on real-world CRV data. The research was supported by various funding sources.", "section_summary": "The section discusses the article \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\" by Yang et al. The purpose of the article is to propose a method for corporate relative valuation using a heterogeneous multi-modal graph neural network. The significance of heterogeneous multi-modal graph neural networks in the field of corporate relative valuation is highlighted, as they allow for the integration of multiple types of data and the capture of complex relationships between entities. The proposed method differs from existing methods for corporate relative valuation in that it uses a graph neural network to model the relationships between entities and incorporates multiple types of data.", "excerpt_keywords": "1. Personalized question recommender system,\n2. Intelligent job interview,\n3. Talent management,\n4. Structure-aware convolutional neural network,\n5. Deep learning,\n6. Airbnb search,\n7. Agency costs,\n8. Corporate finance,\n9. Takeovers,\n10. Free cash flow."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "8f18d25c-ff85-49d7-a478-11d026ab472f", "node_type": "4", "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "d100d58e1005fb2c658803507071ecabcfe9ff3f9f7a07fc712974f29ac6a067"}, "3": {"node_id": "d50f733c-49de-4759-8e76-2362aa141eca", "node_type": "1", "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "3e03fec6a52f958b897685045865c9f13c30ea1d22ea390b591fe9e25df7f0b9"}}, "hash": "c0026d207c549cd92d80fa862c2498a216cf94807d64491e6614a7dcfddf2c8c", "text": "1041-4347 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TKDE.2021.3080293, IEEE\nTransactions on Knowledge and Data Engineering\n13\nREFERENCES\n[1] C. Qin, H. Zhu, C. Zhu, T. Xu, F. Zhuang, C. Ma, J. Zhang,\nand H. Xiong, \u201cDuerquiz: A personalized question rec-\nommender system for intelligent job interview,\u201d in KDD,\nAnchorage, AK, 2019, pp. 2165\u20132173.\n[2] Y. Sun, F. Zhuang, H. Zhu, X. Song, Q. He, and H. Xiong,\n\u201cThe impact of person-organization \ufb01t on talent man-\nagement: A structure-aware convolutional neural network\napproach,\u201d in KDD, Anchorage, AK, 2019, pp. 1625\u20131633.\n[3] M. Haldar, M. Abdool, P . Ramanathan, T. Xu, S. Yang,\nH. Duan, Q. Zhang, N. Barrow-Williams, B. C. Turnbull,\nB. M. Collins, and T. Legrand, \u201cApplying deep learning to\nairbnb search,\u201d in KDD, Anchorage, AK, 2019, pp. 1927\u2013\n1935.\n[4] M. C. Jensen, \u201cAgency costs of free cash \ufb02ow, corporate\n\ufb01nance, and takeovers,\u201d The American economic review,\nvol. 76, no. 2, pp. 323\u2013329, 1986.\n[5] J.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "d50f733c-49de-4759-8e76-2362aa141eca": {"__data__": {"id_": "d50f733c-49de-4759-8e76-2362aa141eca", "embedding": null, "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. How can a heterogeneous multi-modal graph neural network be used for corporate relative valuation?\n2. What are the key factors that influence startup valuation by venture capitalists?\n3. How can attention mechanisms be used in deep learning models for financial analysis?", "prev_section_summary": "The section discusses the article \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\" by Yang et al. The purpose of the article is to propose a method for corporate relative valuation using a heterogeneous multi-modal graph neural network. The significance of heterogeneous multi-modal graph neural networks in the field of corporate relative valuation is highlighted, as they allow for the integration of multiple types of data and the capture of complex relationships between entities. The proposed method differs from existing methods for corporate relative valuation in that it uses a graph neural network to model the relationships between entities and incorporates multiple types of data.", "section_summary": "The section discusses the use of a heterogeneous multi-modal graph neural network for corporate relative valuation, as well as the key factors that influence startup valuation by venture capitalists. The section also explores the use of attention mechanisms in deep learning models for financial analysis. The authors of the section include Zhang, Barrow-Williams, Turnbull, Collins, and Legrand, as well as Jensen, Francis, Luehrman, Fisher, Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser, and Polosukhin. The section also references several studies and books on the topics of corporate finance, startup valuation, and deep learning for financial analysis.", "excerpt_keywords": "deep learning, Airbnb search, KDD, N. Barrow-Williams, B. C. Turnbull, B. M. Collins, T. Legrand, agency costs, free cash flow, corporate finance, takeovers, value-based management, real options, investment opportunities, only three questions that count, investing, knowing what others don't, attention is all you need, startup valuation, venture capitalists, empirical study, determinants of startup valuation, systematic review, avenues for future research."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "8f18d25c-ff85-49d7-a478-11d026ab472f", "node_type": "4", "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "d100d58e1005fb2c658803507071ecabcfe9ff3f9f7a07fc712974f29ac6a067"}, "2": {"node_id": "b163aa80-b9e0-45d9-ac32-f3d3df8ff57e", "node_type": "1", "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "c0026d207c549cd92d80fa862c2498a216cf94807d64491e6614a7dcfddf2c8c"}, "3": {"node_id": "19022137-33ca-4bc4-9c25-84583e635017", "node_type": "1", "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "91437bbe0b922a5587814739b7a93aa6708472eaaa484db7e2789cd806ec3451"}}, "hash": "3e03fec6a52f958b897685045865c9f13c30ea1d22ea390b591fe9e25df7f0b9", "text": "Q. Zhang, N. Barrow-Williams, B. C. Turnbull,\nB. M. Collins, and T. Legrand, \u201cApplying deep learning to\nairbnb search,\u201d in KDD, Anchorage, AK, 2019, pp. 1927\u2013\n1935.\n[4] M. C. Jensen, \u201cAgency costs of free cash \ufb02ow, corporate\n\ufb01nance, and takeovers,\u201d The American economic review,\nvol. 76, no. 2, pp. 323\u2013329, 1986.\n[5] J. Francis, \u201cEva [r] and value-based management: a prac-\ntical guide to implementation.(book reviews),\u201d Accounting\nReview, vol. 77, no. 1, pp. 228\u2013229, 2002.\n[6] T. A. Luehrman, Investment opportunities as real options:\nGetting started on the numbers. Harvard Business Review\nBoston, 1998.\n[7] K. L. Fisher, The only three questions that count: investing by\nknowing what others don\u2019t. John Wiley & Sons, 2007, vol. 22.\n[8] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones,\nA. N. Gomez, L. Kaiser, and I. Polosukhin, \u201cAttention is all\nyou need,\u201d in NeurIPS, Long Beach, CA, 2017, pp. 5998\u2013\n6008.\n[9] T. Miloud, A. Aspelund, and M. Cabrol, \u201cStartup valuation\nby venture capitalists: an empirical study,\u201d Venture Capital,\nvol. 14, no. 2-3, pp. 151\u2013174, 2012.\n[10] A. Kohn, \u201cThe determinants of startup valuation in the\nventure capital context: a systematic review and avenues\nfor future research,\u201d Management Review Quarterly, vol. 68,\nno. 1, pp. 3\u201336, 2018.\n[11]", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "19022137-33ca-4bc4-9c25-84583e635017": {"__data__": {"id_": "19022137-33ca-4bc4-9c25-84583e635017", "embedding": null, "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the relationship between startup valuation and academic social networks?\n2. How can heterogeneous graph neural networks be used to model the visual evolution of fashion trends?\n3. What are the key features of heterogeneous graph neural networks and how do they differ from other graph neural networks?", "prev_section_summary": "The section discusses the use of a heterogeneous multi-modal graph neural network for corporate relative valuation, as well as the key factors that influence startup valuation by venture capitalists. The section also explores the use of attention mechanisms in deep learning models for financial analysis. The authors of the section include Zhang, Barrow-Williams, Turnbull, Collins, and Legrand, as well as Jensen, Francis, Luehrman, Fisher, Vaswani, Shazeer, Parmar, Uszkoreit, Jones, Gomez, Kaiser, and Polosukhin. The section also references several studies and books on the topics of corporate finance, startup valuation, and deep learning for financial analysis.", "section_summary": "The section discusses the relationship between startup valuation and academic social networks, as well as the use of heterogeneous graph neural networks to model the visual evolution of fashion trends. The key features of heterogeneous graph neural networks are also discussed, along with their differences from other graph neural networks. The section includes references to several studies and papers on these topics, including [9], [10], [11], [12], [13], [14], and [15].", "excerpt_keywords": "startup valuation, venture capital, empirical study, determinants, management review quarterly, systematic review, avenues for future research, academic social networks, fashion trends, visual evolution, one-class collaborative filtering, heterogeneous graph neural network, node2vec, scalable feature learning, networks, heterogeneous entity graphs, linking large-scale heterogeneous entity graphs."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "8f18d25c-ff85-49d7-a478-11d026ab472f", "node_type": "4", "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "d100d58e1005fb2c658803507071ecabcfe9ff3f9f7a07fc712974f29ac6a067"}, "2": {"node_id": "d50f733c-49de-4759-8e76-2362aa141eca", "node_type": "1", "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "3e03fec6a52f958b897685045865c9f13c30ea1d22ea390b591fe9e25df7f0b9"}, "3": {"node_id": "1cbbd6e3-e7ce-47d5-a560-d15fca7ecc58", "node_type": "1", "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "098d898699ce3eabc62fb7e6ca223d07baf98ec392f4cc447b4abd452ae7cc90"}}, "hash": "91437bbe0b922a5587814739b7a93aa6708472eaaa484db7e2789cd806ec3451", "text": "5998\u2013\n6008.\n[9] T. Miloud, A. Aspelund, and M. Cabrol, \u201cStartup valuation\nby venture capitalists: an empirical study,\u201d Venture Capital,\nvol. 14, no. 2-3, pp. 151\u2013174, 2012.\n[10] A. Kohn, \u201cThe determinants of startup valuation in the\nventure capital context: a systematic review and avenues\nfor future research,\u201d Management Review Quarterly, vol. 68,\nno. 1, pp. 3\u201336, 2018.\n[11] J. Tang, J. Zhang, L. Yao, J. Li, L. Zhang, and Z. Su,\n\u201cArnetminer: extraction and mining of academic social\nnetworks,\u201d in KDD, Las Vegas, Nevada, 2008, pp. 990\u2013998.\n[12] R. He and J. J. McAuley, \u201cUps and downs: Modeling the\nvisual evolution of fashion trends with one-class collabo-\nrative \ufb01ltering,\u201d in WWW, Montreal, Canada, pp. 507\u2013517.\n[13] C. Zhang, D. Song, C. Huang, A. Swami, and N. V .\nChawla, \u201cHeterogeneous graph neural network,\u201d in KDD,\nAnchorage, AK, 2019, pp. 793\u2013803.\n[14] A. Grover and J. Leskovec, \u201cnode2vec: Scalable feature\nlearning for networks,\u201d in KDD, San Francisco, CA, 2016,\npp. 855\u2013864.\n[15] F. Zhang, X. Liu, J. Tang, Y. Dong, P . Yao, J. Zhang, X. Gu,\nY. Wang, B. Shao, R. Li, and K. Wang, \u201cOAG: toward\nlinking large-scale heterogeneous entity graphs,\u201d in KDD,\nAnchorage, AK, 2019, pp.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "1cbbd6e3-e7ce-47d5-a560-d15fca7ecc58": {"__data__": {"id_": "1cbbd6e3-e7ce-47d5-a560-d15fca7ecc58", "embedding": null, "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using heterogeneous multi-modal graph neural networks for corporate relative valuation?\n2. How does the use of heterogeneous multi-modal graph neural networks improve the accuracy of corporate relative valuation?\n3. What are some of the limitations or challenges associated with using heterogeneous multi-modal graph neural networks for corporate relative valuation?", "prev_section_summary": "The section discusses the relationship between startup valuation and academic social networks, as well as the use of heterogeneous graph neural networks to model the visual evolution of fashion trends. The key features of heterogeneous graph neural networks are also discussed, along with their differences from other graph neural networks. The section includes references to several studies and papers on these topics, including [9], [10], [11], [12], [13], [14], and [15].", "section_summary": "The section discusses the use of heterogeneous multi-modal graph neural networks for corporate relative valuation. The purpose of using such networks is to improve the accuracy of valuation by incorporating multiple types of data and relationships. The section mentions several papers and techniques related to graph neural networks and their applications, including node2vec, OAG, IntentGC, long short-term memory, inductive representation learning, heterogeneous graph neural networks for malicious account detection, and semi-supervised classification with graph convolutional networks.", "excerpt_keywords": "1. Graph neural networks\n2. Heterogeneous information\n3. Recommendation systems\n4. Linking large-scale entity graphs\n5. Scalable feature learning\n6. Long short-term memory\n7. Inductive representation learning\n8. Malicious account detection\n9. Semi-supervised classification\n10. Graph convolutional networks"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "8f18d25c-ff85-49d7-a478-11d026ab472f", "node_type": "4", "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "d100d58e1005fb2c658803507071ecabcfe9ff3f9f7a07fc712974f29ac6a067"}, "2": {"node_id": "19022137-33ca-4bc4-9c25-84583e635017", "node_type": "1", "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "91437bbe0b922a5587814739b7a93aa6708472eaaa484db7e2789cd806ec3451"}, "3": {"node_id": "c81e1197-5c10-4242-9505-acf800451254", "node_type": "1", "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "95e564e82179d3613ea61dd07d9dc8ea8c647ce9020fd7344a687bd4b264681f"}}, "hash": "098d898699ce3eabc62fb7e6ca223d07baf98ec392f4cc447b4abd452ae7cc90", "text": "and J. Leskovec, \u201cnode2vec: Scalable feature\nlearning for networks,\u201d in KDD, San Francisco, CA, 2016,\npp. 855\u2013864.\n[15] F. Zhang, X. Liu, J. Tang, Y. Dong, P . Yao, J. Zhang, X. Gu,\nY. Wang, B. Shao, R. Li, and K. Wang, \u201cOAG: toward\nlinking large-scale heterogeneous entity graphs,\u201d in KDD,\nAnchorage, AK, 2019, pp. 2585\u20132595.\n[16] J. Zhao, Z. Zhou, Z. Guan, W. Zhao, W. Ning, G. Qiu, and\nX. He, \u201cIntentgc: A scalable graph convolution framework\nfusing heterogeneous information for recommendation,\u201d\ninKDD, Anchorage, AK, 2019, pp. 2347\u20132357.\n[17] S. Hochreiter and J. Schmidhuber, \u201cLong short-term mem-\nory,\u201d Neural Computation, vol. 9, no. 8, pp. 1735\u20131780, 1997.\n[18] W. L. Hamilton, Z. Ying, and J. Leskovec, \u201cInductive\nrepresentation learning on large graphs,\u201d in NeurIPS, Long\nBeach, CA, 2017, pp. 1024\u20131034.\n[19] Z. Liu, C. Chen, X. Yang, J. Zhou, X. Li, and L. Song, \u201cHet-\nerogeneous graph neural networks for malicious account\ndetection,\u201d in CIKM, Torino, Italy, 2018, pp. 2077\u20132085.\n[20] T. N. Kipf and M. Welling, \u201cSemi-supervised classi\ufb01cation\nwith graph convolutional networks,\u201d in ICLR, Toulon,\nFrance, 2017.\n[21] E. B.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "c81e1197-5c10-4242-9505-acf800451254": {"__data__": {"id_": "c81e1197-5c10-4242-9505-acf800451254", "embedding": null, "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of heterogeneous graph neural networks in the field of malicious account detection?\n2. How do graph convolutional networks contribute to semi-supervised classification?\n3. What are some applications of graph neural networks beyond malicious account detection and semi-supervised classification?", "prev_section_summary": "The section discusses the use of heterogeneous multi-modal graph neural networks for corporate relative valuation. The purpose of using such networks is to improve the accuracy of valuation by incorporating multiple types of data and relationships. The section mentions several papers and techniques related to graph neural networks and their applications, including node2vec, OAG, IntentGC, long short-term memory, inductive representation learning, heterogeneous graph neural networks for malicious account detection, and semi-supervised classification with graph convolutional networks.", "section_summary": "The section discusses the use of heterogeneous graph neural networks in various applications, including malicious account detection and semi-supervised classification. The authors of the section provide references to several papers that explore the use of graph neural networks in these applications. The section also mentions some other applications of graph neural networks beyond these two areas.", "excerpt_keywords": "graph neural networks, malicious account detection, semi-supervised classification, combinatorial optimization algorithms, graph convolutional networks, graph representation learning, word representations, efficient estimation, feature propagation, graph attention networks."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "8f18d25c-ff85-49d7-a478-11d026ab472f", "node_type": "4", "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "d100d58e1005fb2c658803507071ecabcfe9ff3f9f7a07fc712974f29ac6a067"}, "2": {"node_id": "1cbbd6e3-e7ce-47d5-a560-d15fca7ecc58", "node_type": "1", "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "098d898699ce3eabc62fb7e6ca223d07baf98ec392f4cc447b4abd452ae7cc90"}, "3": {"node_id": "6283fc3c-7275-4b4e-8d9b-50df3aebb288", "node_type": "1", "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "c1f258daf89ca0070248f51f0809cd3a059392a1a78d6280b6863b8c8a9c080c"}}, "hash": "95e564e82179d3613ea61dd07d9dc8ea8c647ce9020fd7344a687bd4b264681f", "text": "Z. Liu, C. Chen, X. Yang, J. Zhou, X. Li, and L. Song, \u201cHet-\nerogeneous graph neural networks for malicious account\ndetection,\u201d in CIKM, Torino, Italy, 2018, pp. 2077\u20132085.\n[20] T. N. Kipf and M. Welling, \u201cSemi-supervised classi\ufb01cation\nwith graph convolutional networks,\u201d in ICLR, Toulon,\nFrance, 2017.\n[21] E. B. Khalil, H. Dai, Y. Zhang, B. Dilkina, and L. Song,\n\u201cLearning combinatorial optimization algorithms overgraphs,\u201d in NeurIPS, Long Beach, CA, 2017, pp. 6348\u20136358.\n[22] J. Zhou, G. Cui, Z. Zhang, C. Yang, Z. Liu, and M. Sun,\n\u201cGraph neural networks: A review of methods and appli-\ncations,\u201d CoRR, vol. abs/1812.08434, 2018.\n[23] T. Mikolov, K. Chen, G. Corrado, and J. Dean, \u201cEf\ufb01cient\nestimation of word representations in vector space,\u201d in\nICLR Workshop Track, Scottsdale, Arizona, 2013.\n[24] B. Xiang, Z. Liu, J. Zhou, and X. Li, \u201cFeature propagation\non graph: A new perspective to graph representation\nlearning,\u201d CoRR, vol. abs/1804.06111, 2018.\n[25] P . Velickovic, G. Cucurull, A. Casanova, A. Romero, P . Lio,\nand Y. Bengio, \u201cGraph attention networks,\u201d in ICLR, Van-\ncouver, BC, 2018.\n[26] B. Perozzi, R. Al-Rfou, and S. Skiena,", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "6283fc3c-7275-4b4e-8d9b-50df3aebb288": {"__data__": {"id_": "6283fc3c-7275-4b4e-8d9b-50df3aebb288", "embedding": null, "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the paper \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\"?\n2. What are the key contributions of the paper \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\"?\n3. How does the paper \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\" compare to other graph neural network models for corporate relative valuation?", "prev_section_summary": "The section discusses the use of heterogeneous graph neural networks in various applications, including malicious account detection and semi-supervised classification. The authors of the section provide references to several papers that explore the use of graph neural networks in these applications. The section also mentions some other applications of graph neural networks beyond these two areas.", "section_summary": "The section discusses the paper \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\" by Yang et al. The paper proposes a graph neural network model for corporate relative valuation that incorporates multiple modalities of data. The key contributions of the paper include the use of heterogeneous multi-modal graph neural networks, the incorporation of multiple modalities of data, and the ability to handle complex relationships between entities. The paper compares its model to other graph neural network models for corporate relative valuation and discusses its advantages and limitations. The section also mentions several other papers related to graph neural network models for corporate relative valuation and representation learning for heterogeneous networks.", "excerpt_keywords": "graph representation learning, feature propagation, graph attention networks, deepwalk, Adam, metapath2vec, attributed social network embedding, heterogeneous graph attention network, representation learning for attributed multiplex heterogeneous network, graph neural networks."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "8f18d25c-ff85-49d7-a478-11d026ab472f", "node_type": "4", "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "d100d58e1005fb2c658803507071ecabcfe9ff3f9f7a07fc712974f29ac6a067"}, "2": {"node_id": "c81e1197-5c10-4242-9505-acf800451254", "node_type": "1", "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "95e564e82179d3613ea61dd07d9dc8ea8c647ce9020fd7344a687bd4b264681f"}, "3": {"node_id": "57e041d4-db81-42b9-b181-900c2666dba8", "node_type": "1", "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "e8cf7caa3ce3516a335ace2a7c7c0a688b32650f276cd2dacc2af1af228fa134"}}, "hash": "c1f258daf89ca0070248f51f0809cd3a059392a1a78d6280b6863b8c8a9c080c", "text": "Zhou, and X. Li, \u201cFeature propagation\non graph: A new perspective to graph representation\nlearning,\u201d CoRR, vol. abs/1804.06111, 2018.\n[25] P . Velickovic, G. Cucurull, A. Casanova, A. Romero, P . Lio,\nand Y. Bengio, \u201cGraph attention networks,\u201d in ICLR, Van-\ncouver, BC, 2018.\n[26] B. Perozzi, R. Al-Rfou, and S. Skiena, \u201cDeepwalk: online\nlearning of social representations,\u201d in KDD, New York, NY,\n2014, pp. 701\u2013710.\n[27] D. P . Kingma and J. Ba, \u201cAdam: A method for stochastic\noptimization,\u201d in ICLR, San Diego, CA, 2015.\n[28] Y. Dong, N. V . Chawla, and A. Swami, \u201cmetapath2vec:\nScalable representation learning for heterogeneous net-\nworks,\u201d in KDD, Halifax, Canada, 2017, pp. 135\u2013144.\n[29] L. Liao, X. He, H. Zhang, and T. Chua, \u201cAttributed social\nnetwork embedding,\u201d TKDE, vol. 30, no. 12, pp. 2257\u2013\n2270, 2018.\n[30] X. Wang, H. Ji, C. Shi, B. Wang, Y. Ye, P . Cui, and P . S. Yu,\n\u201cHeterogeneous graph attention network,\u201d in WWW, San\nFrancisco, CA, 2019, pp. 2022\u20132032.\n[31] Y. Cen, X. Zou, J. Zhang, H. Yang, J. Zhou, and J. Tang,\n\u201cRepresentation learning for attributed multiplex hetero-\ngeneous network,\u201d in KDD, Anchorage, AK, 2019,", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "57e041d4-db81-42b9-b181-900c2666dba8": {"__data__": {"id_": "57e041d4-db81-42b9-b181-900c2666dba8", "embedding": null, "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the relationship between heterogeneous multi-modal graph neural networks and corporate relative valuation?\n2. How does the use of graph attention networks improve the accuracy of corporate relative valuation?\n3. What are the limitations of using price-to-earnings ratio and price-to-book ratio as valuation models for corporations?", "prev_section_summary": "The section discusses the paper \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\" by Yang et al. The paper proposes a graph neural network model for corporate relative valuation that incorporates multiple modalities of data. The key contributions of the paper include the use of heterogeneous multi-modal graph neural networks, the incorporation of multiple modalities of data, and the ability to handle complex relationships between entities. The paper compares its model to other graph neural network models for corporate relative valuation and discusses its advantages and limitations. The section also mentions several other papers related to graph neural network models for corporate relative valuation and representation learning for heterogeneous networks.", "section_summary": "The section discusses the use of heterogeneous multi-modal graph neural networks for corporate relative valuation. The authors of the section propose a graph attention network that can improve the accuracy of corporate relative valuation. They also discuss the limitations of using price-to-earnings ratio and price-to-book ratio as valuation models for corporations. The section also includes references to other research on the topic, including a review of the price-to-earnings ratio and an empirical investigation of the price-to-book ratio as a valuation model.", "excerpt_keywords": "heterogeneous graph attention network, representation learning, attributed multiplex heterogeneous network, price-to-earnings ratio, price-to-book ratio, valuation model, investment strategies, real estate valuation, artificial intelligence, corporate governance, equity valuation, neural network."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "8f18d25c-ff85-49d7-a478-11d026ab472f", "node_type": "4", "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "d100d58e1005fb2c658803507071ecabcfe9ff3f9f7a07fc712974f29ac6a067"}, "2": {"node_id": "6283fc3c-7275-4b4e-8d9b-50df3aebb288", "node_type": "1", "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "c1f258daf89ca0070248f51f0809cd3a059392a1a78d6280b6863b8c8a9c080c"}, "3": {"node_id": "f5ded79c-9fee-4f07-8065-576f28712bea", "node_type": "1", "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "fdfc4783d4172c9d5956dfc26ff42d1c830acde8b1f0aefbeb1f5031e686b296"}}, "hash": "e8cf7caa3ce3516a335ace2a7c7c0a688b32650f276cd2dacc2af1af228fa134", "text": "X. Wang, H. Ji, C. Shi, B. Wang, Y. Ye, P . Cui, and P . S. Yu,\n\u201cHeterogeneous graph attention network,\u201d in WWW, San\nFrancisco, CA, 2019, pp. 2022\u20132032.\n[31] Y. Cen, X. Zou, J. Zhang, H. Yang, J. Zhou, and J. Tang,\n\u201cRepresentation learning for attributed multiplex hetero-\ngeneous network,\u201d in KDD, Anchorage, AK, 2019, pp.\n1358\u20131368.\n[32] M. Ghaeli, \u201cPrice-to-earnings ratio: A state-of-art review,\u201d\nAccounting, vol. 3, no. 2, pp. 131\u2013136, 2017.\n[33] S. Agrawal, R. Monem, and M. Ariff, \u201cPrice to book ratio\nas a valuation model: An empirical investigation,\u201d Finance\nIndia, vol. 2, no. 10, pp. 333\u2013344, 1996.\n[34] J. P . O\u2019Shaughnessy, What works on Wall Street: A guide to the\nbest-performing investment strategies of all time. McGraw-\nHill, 1998.\n[35] E. Pagourtzi, K. Metaxiotis, K. Nikolopoulos, K. Gianne-\nlos, and V . Assimakopoulos, \u201cReal estate valuation with\narti\ufb01cial intelligence approaches,\u201d ISTA, vol. 2, no. 1, pp.\n50\u201357, 2007.\n[36] G. Karami and S. BeikBoshrouyeh, \u201cCorporate governance\nand equity valuation: the model by using arti\ufb01cial neu-\nral network,\u201d Journal of Accounting and Auditing Review,\nvol. 18, no. 64, pp. 129\u2013150, 2011.\n[37] G. Evangelopoulos, A. Zlatintsi,", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "f5ded79c-9fee-4f07-8065-576f28712bea": {"__data__": {"id_": "f5ded79c-9fee-4f07-8065-576f28712bea", "embedding": null, "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the relationship between heterogeneous multi-modal graph neural networks and corporate relative valuation?\n2. How can multi-modal saliency and fusion be used to improve the accuracy of real estate valuation with artificial intelligence approaches?\n3. What are some examples of applications of multi-modal fusion with recurrent neural networks in the field of natural language processing?", "prev_section_summary": "The section discusses the use of heterogeneous multi-modal graph neural networks for corporate relative valuation. The authors of the section propose a graph attention network that can improve the accuracy of corporate relative valuation. They also discuss the limitations of using price-to-earnings ratio and price-to-book ratio as valuation models for corporations. The section also includes references to other research on the topic, including a review of the price-to-earnings ratio and an empirical investigation of the price-to-book ratio as a valuation model.", "section_summary": "The section discusses the use of heterogeneous multi-modal graph neural networks in the field of corporate relative valuation. The authors propose a model that uses multi-modal saliency and fusion to improve the accuracy of real estate valuation with artificial intelligence approaches. The section also provides examples of applications of multi-modal fusion with recurrent neural networks in natural language processing, such as movie summarization and rumor detection on microblogs. The authors also mention the use of multiple kernel learning algorithms and comprehensive semi-supervised multi-modal learning in other fields.", "excerpt_keywords": "Artificial intelligence, real estate valuation, corporate governance, equity valuation, multi-modal saliency, fusion, movie summarization, multiple kernel learning, rumor detection, microblogs, comprehensive semi-supervised multi-modal learning, network embedding."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "8f18d25c-ff85-49d7-a478-11d026ab472f", "node_type": "4", "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "d100d58e1005fb2c658803507071ecabcfe9ff3f9f7a07fc712974f29ac6a067"}, "2": {"node_id": "57e041d4-db81-42b9-b181-900c2666dba8", "node_type": "1", "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "e8cf7caa3ce3516a335ace2a7c7c0a688b32650f276cd2dacc2af1af228fa134"}, "3": {"node_id": "062d9cab-f298-462f-af4b-0c2cfb52a456", "node_type": "1", "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "31dbf21418eec94113cb6b19f6925b9c2e20f611b426f47004fa03566e45c0ad"}}, "hash": "fdfc4783d4172c9d5956dfc26ff42d1c830acde8b1f0aefbeb1f5031e686b296", "text": "\u201cReal estate valuation with\narti\ufb01cial intelligence approaches,\u201d ISTA, vol. 2, no. 1, pp.\n50\u201357, 2007.\n[36] G. Karami and S. BeikBoshrouyeh, \u201cCorporate governance\nand equity valuation: the model by using arti\ufb01cial neu-\nral network,\u201d Journal of Accounting and Auditing Review,\nvol. 18, no. 64, pp. 129\u2013150, 2011.\n[37] G. Evangelopoulos, A. Zlatintsi, A. Potamianos, P . Mara-\ngos, K. Rapantzikos, G. Skoumas, and Y. Avrithis, \u201cMulti-\nmodal saliency and fusion for movie summarization based\non aural, visual, and textual attention,\u201d TMM, vol. 15,\nno. 7, pp. 1553\u20131568, 2013.\n[38] M. Gonen and E. Alpaydin, \u201cMultiple kernel learning\nalgorithms,\u201d JMLR, vol. 12, pp. 2211\u20132268, 2011.\n[39] Z. Jin, J. Cao, H. Guo, Y. Zhang, and J. Luo, \u201cMultimodal\nfusion with recurrent neural networks for rumor detection\non microblogs,\u201d in ACMMMM, Mountain View, CA, 2017,\npp. 795\u2013816.\n[40] Y. Yang, K. Wang, D. Zhan, H. Xiong, and Y. Jiang,\n\u201cComprehensive semi-supervised multi-modal learning,\u201d\ninIJCAI, Macao, China, 2019, pp. 4092\u20134098.\n[41] P . Cui, X. Wang, J. Pei, and W. Zhu, \u201cA survey on network\nembedding,\u201d TKDE, vol. 31, no. 5, pp. 833\u2013852,", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "062d9cab-f298-462f-af4b-0c2cfb52a456": {"__data__": {"id_": "062d9cab-f298-462f-af4b-0c2cfb52a456", "embedding": null, "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the difference between a heterogeneous multi-modal graph neural network and a traditional graph neural network?\n2. How can a semi-supervised multi-modal learning approach be used for corporate relative valuation?\n3. What is the role of meta-path based context in top-N recommendation with a neural co-attention model?", "prev_section_summary": "The section discusses the use of heterogeneous multi-modal graph neural networks in the field of corporate relative valuation. The authors propose a model that uses multi-modal saliency and fusion to improve the accuracy of real estate valuation with artificial intelligence approaches. The section also provides examples of applications of multi-modal fusion with recurrent neural networks in natural language processing, such as movie summarization and rumor detection on microblogs. The authors also mention the use of multiple kernel learning algorithms and comprehensive semi-supervised multi-modal learning in other fields.", "section_summary": "The section discusses the use of heterogeneous multi-modal graph neural networks for corporate relative valuation. The key topics include the difference between heterogeneous multi-modal graph neural networks and traditional graph neural networks, the role of meta-path based context in top-N recommendation with a neural co-attention model, and the use of semi-supervised multi-modal learning for corporate relative valuation. The entities mentioned in the section include Y. Yang, K. Wang, D. Zhan, H. Xiong, and Y. Jiang, P. Cui, X. Wang, J. Pei, and W. Zhu, Y. Sun, J. Han, C. C. Aggarwal, and N. V. Chawla, B. Hu, C. Shi, W. X. Zhao, and P. S. Yu, and IEEE Xplore.", "excerpt_keywords": "1. Semi-supervised learning\n2. Multi-modal learning\n3. Network embedding\n4. Relationship prediction\n5. Heterogeneous information networks\n6. Meta-path based context\n7. Top-N recommendation\n8. Neural co-attention model\n9. Relevance feedback\n10. Personalized recommendation"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "8f18d25c-ff85-49d7-a478-11d026ab472f", "node_type": "4", "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "d100d58e1005fb2c658803507071ecabcfe9ff3f9f7a07fc712974f29ac6a067"}, "2": {"node_id": "f5ded79c-9fee-4f07-8065-576f28712bea", "node_type": "1", "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "fdfc4783d4172c9d5956dfc26ff42d1c830acde8b1f0aefbeb1f5031e686b296"}}, "hash": "31dbf21418eec94113cb6b19f6925b9c2e20f611b426f47004fa03566e45c0ad", "text": "795\u2013816.\n[40] Y. Yang, K. Wang, D. Zhan, H. Xiong, and Y. Jiang,\n\u201cComprehensive semi-supervised multi-modal learning,\u201d\ninIJCAI, Macao, China, 2019, pp. 4092\u20134098.\n[41] P . Cui, X. Wang, J. Pei, and W. Zhu, \u201cA survey on network\nembedding,\u201d TKDE, vol. 31, no. 5, pp. 833\u2013852, 2019.\n[42] Y. Sun, J. Han, C. C. Aggarwal, and N. V . Chawla, \u201cWhen\nwill it happen?: relationship prediction in heterogeneous\ninformation networks,\u201d in WWW, Seattle, WA, 2012, pp.\n663\u2013672.\n[43] B. Hu, C. Shi, W. X. Zhao, and P . S. Yu, \u201cLeveraging meta-\npath based context for top- N recommendation with A\nneural co-attention model,\u201d in KDD, London, UK, 2018,\nAuthorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 20,2021 at 08:36:01 UTC from IEEE Xplore.  Restrictions apply.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "39e4ff77-a4f5-498e-a2fb-94f577ae7908": {"__data__": {"id_": "39e4ff77-a4f5-498e-a2fb-94f577ae7908", "embedding": null, "metadata": {"page_label": "14", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the title of the article and who are the authors?\n2. What is the journal that the article has been accepted for publication in?\n3. What is the research focus of the authors and what are some of their previous publications in this area?", "prev_section_summary": "The section discusses the use of heterogeneous multi-modal graph neural networks for corporate relative valuation. The key topics include the difference between heterogeneous multi-modal graph neural networks and traditional graph neural networks, the role of meta-path based context in top-N recommendation with a neural co-attention model, and the use of semi-supervised multi-modal learning for corporate relative valuation. The entities mentioned in the section include Y. Yang, K. Wang, D. Zhan, H. Xiong, and Y. Jiang, P. Cui, X. Wang, J. Pei, and W. Zhu, Y. Sun, J. Han, C. C. Aggarwal, and N. V. Chawla, B. Hu, C. Shi, W. X. Zhao, and P. S. Yu, and IEEE Xplore.", "section_summary": "The section discusses a research article titled \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\" by Yang et al. The article has been accepted for publication in the journal Transactions on Knowledge and Data Engineering. The authors' research interests include machine learning and data mining, specifically heterogeneous learning, model reuse, and incremental mining. The authors have published over 10 papers in leading international journals and conferences, and serve as PCs in conferences such as IJCAI, AAAI, ICML, and NIPS. The article focuses on the use of heterogeneous multi-modal graph neural networks for corporate relative valuation.", "excerpt_keywords": "1. Heterogeneous network embedding,\n2. Attributed multiplex heterogeneous network embedding,\n3. Multi-source information fusion,\n4. Machine learning,\n5. Data mining,\n6. Incremental mining,\n7. Model reuse,\n8. Novel software technology,\n9. National Key Lab,\n10. Nanjing University."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "b1879836-b599-41e1-af94-4bfe72a77860", "node_type": "4", "metadata": {"page_label": "14", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a8f60bb292a6a8a5a8797722564372a56ffb6de873685c2bc9b524a068f8aafa"}, "3": {"node_id": "a4b15fbe-d158-4265-b5ef-7acf0309dbb3", "node_type": "1", "metadata": {"page_label": "14", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "6d0c1f0fa5ef555549109e642e9a987e4eaf08ec2941df7e7ef9575b54dbc56b"}}, "hash": "83e8e42125e4c477446b67f8b03fe3a6b437d870c4b11109505de14fb0b823bc", "text": "1041-4347 (c) 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TKDE.2021.3080293, IEEE\nTransactions on Knowledge and Data Engineering\n14\npp. 1531\u20131540.\n[44] Z. Liu, C. Huang, Y. Yu, B. Fan, and J. Dong, \u201cFast\nattributed multiplex heterogeneous network embedding,\u201d\ninCIKM, Virtual Event, Ireland, 2020, pp. 995\u20131004.\n[45] B. Li, D. Pi, Y. Lin, I. A. Khan, and L. Cui, \u201cMulti-\nsource information fusion based heterogeneous network\nembedding,\u201d Inf. Sci., vol. 534, pp. 53\u201371, 2020.\nYang Yang received the Ph.D. degree in com-\nputer science, Nanjing University, China in 2019.\nAt the same year, he became a faculty member\nat Nanjing University of Science and Technol-\nogy, China. He is currently a Professor with the\nSchool of Computer Science and Engineering.\nHis research interests lie primarily in machine\nlearning and data mining, including heteroge-\nneous learning, model reuse, and incremental\nmining. He has published over 10 papers in lead-\ning international journal/conferences. He serves\nas PC in leading conferences such as IJCAI, AAAI, ICML, NIPS, etc.\nJia-Qi Yang is working towards the M.Sc. de-\ngree with the National Key Lab for Novel Soft-\nware Technology, the Department of Computer\nScience &Technology in Nanjing University,\nChina. His research interests lie primarily in ma-\nchine learning and", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "a4b15fbe-d158-4265-b5ef-7acf0309dbb3": {"__data__": {"id_": "a4b15fbe-d158-4265-b5ef-7acf0309dbb3", "embedding": null, "metadata": {"page_label": "14", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the research focus of Jia-Qi Yang, Ran Bao, and De-Chuan Zhan in the field of machine learning and data mining?\n2. What are the specific research interests of HengShu Zhu in the field of data mining and machine learning?\n3. What are the notable publications and conferences that Jia-Qi Yang, Ran Bao, De-Chuan Zhan, and HengShu Zhu have contributed to in the field of machine learning and data mining?", "prev_section_summary": "The section discusses a research article titled \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\" by Yang et al. The article has been accepted for publication in the journal Transactions on Knowledge and Data Engineering. The authors' research interests include machine learning and data mining, specifically heterogeneous learning, model reuse, and incremental mining. The authors have published over 10 papers in leading international journals and conferences, and serve as PCs in conferences such as IJCAI, AAAI, ICML, and NIPS. The article focuses on the use of heterogeneous multi-modal graph neural networks for corporate relative valuation.", "section_summary": "The section discusses the research focus and interests of Jia-Qi Yang, Ran Bao, De-Chuan Zhan, and HengShu Zhu in the field of machine learning and data mining. Jia-Qi Yang's research interests include multi-modal learning, while Ran Bao's research interests lie primarily in machine learning and data mining. De-Chuan Zhan's research interests are mainly in machine learning, data mining, and mobile intelligence. HengShu Zhu's research focuses on developing advanced data analysis techniques. The section also mentions notable publications and conferences that these researchers have contributed to, as well as their affiliations and service roles in leading conferences.", "excerpt_keywords": "Machine learning, data mining, multi-modal learning, heterogeneous learning, model reuse, incremental mining, National Key Lab for Novel Software Technology, Nanjing University, China, M.Sc. degree, Ph.D. degree, Professor, editorial board member, SPC/PC, IDA, IJAPR, IJCAI, AAAI, ICML, NIPS, Baidu Inc., principal data scientist, architecture."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "b1879836-b599-41e1-af94-4bfe72a77860", "node_type": "4", "metadata": {"page_label": "14", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a8f60bb292a6a8a5a8797722564372a56ffb6de873685c2bc9b524a068f8aafa"}, "2": {"node_id": "39e4ff77-a4f5-498e-a2fb-94f577ae7908", "node_type": "1", "metadata": {"page_label": "14", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "83e8e42125e4c477446b67f8b03fe3a6b437d870c4b11109505de14fb0b823bc"}, "3": {"node_id": "856e13fe-3def-4ed6-b440-d474562d6790", "node_type": "1", "metadata": {"page_label": "14", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "2e5bb07417d7a1c0f9a9fec667832144d3ccf5f14121a46cb977eec81ed7066a"}}, "hash": "6d0c1f0fa5ef555549109e642e9a987e4eaf08ec2941df7e7ef9575b54dbc56b", "text": "and data mining, including heteroge-\nneous learning, model reuse, and incremental\nmining. He has published over 10 papers in lead-\ning international journal/conferences. He serves\nas PC in leading conferences such as IJCAI, AAAI, ICML, NIPS, etc.\nJia-Qi Yang is working towards the M.Sc. de-\ngree with the National Key Lab for Novel Soft-\nware Technology, the Department of Computer\nScience &Technology in Nanjing University,\nChina. His research interests lie primarily in ma-\nchine learning and data mining, including multi-\nmodal learning.\nRan Bao is working towards the M.Sc. degree\nwith the National Key Lab for Novel Software\nTechnology, the Department of Computer Sci-\nence &Technology in Nanjing University, China.\nHis research interests lie primarily in machine\nlearning and data mining.\nDe-Chuan Zhan received the Ph.D. degree in\ncomputer science, Nanjing University, China in\n2010. At the same year, he became a faculty\nmember in the Department of Computer Science\nand Technology at Nanjing University, China. He\nis currently a Professor with the Department of\nComputer Science and Technology at Nanjing\nUniversity. His research interests are mainly in\nmachine learning, data mining and mobile intelli-\ngence. He has published over 20 papers in lead-\ning international journal/conferences. He serves\nas an editorial board member of IDA and IJAPR, and serves as SPC/PC\nin leading conferences such as IJCAI, AAAI, ICML, NIPS, etc.\nHengShu Zhu (SM\u201919) is currently a principal\ndata scientist &architect at Baidu Inc. He re-\nceived the Ph.D. degree in 2014 and B.E. de-\ngree in 2009, both in Computer Science from\nUniversity of Science and Technology of China\n(USTC), China. His general area of research is\ndata mining and machine learning, with a fo-\ncus on developing advanced data analysis", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "856e13fe-3def-4ed6-b440-d474562d6790": {"__data__": {"id_": "856e13fe-3def-4ed6-b440-d474562d6790", "embedding": null, "metadata": {"page_label": "14", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the research focus of HengShu Zhu, a principal data scientist & architect at Baidu Inc.?\n2. What are some of the conferences that HengShu Zhu has served as an organization or program committee member, and what were his roles in these conferences?\n3. What are some of the awards and recognitions that HengShu Zhu has received for his research and contributions in the field of data mining and machine learning?", "prev_section_summary": "The section discusses the research focus and interests of Jia-Qi Yang, Ran Bao, De-Chuan Zhan, and HengShu Zhu in the field of machine learning and data mining. Jia-Qi Yang's research interests include multi-modal learning, while Ran Bao's research interests lie primarily in machine learning and data mining. De-Chuan Zhan's research interests are mainly in machine learning, data mining, and mobile intelligence. HengShu Zhu's research focuses on developing advanced data analysis techniques. The section also mentions notable publications and conferences that these researchers have contributed to, as well as their affiliations and service roles in leading conferences.", "section_summary": "The section discusses the research focus and contributions of HengShu Zhu, a principal data scientist & architect at Baidu Inc. His research is in the field of data mining and machine learning, with a focus on developing advanced data analysis techniques for innovative business applications. He has published prolifically in refereed journals and conference proceedings, and has served regularly on the organization and program committees of numerous conferences. He has received several awards and recognitions for his research and contributions in the field, including the Distinguished Dissertation Award of CAS and the Best Student Paper Award of several conferences. The section also mentions Xiao-Ru Gao, who is working towards a Ph.D. degree in Marketing.", "excerpt_keywords": "1. Ph.D. degree\n2. Marketing\n3. Doctoral program\n4. Academic research\n5. Scholarship\n6. Best student paper award\n7. Conference proceedings\n8. Refereed journals\n9. Program committee\n10. Organizational behavior"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "b1879836-b599-41e1-af94-4bfe72a77860", "node_type": "4", "metadata": {"page_label": "14", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a8f60bb292a6a8a5a8797722564372a56ffb6de873685c2bc9b524a068f8aafa"}, "2": {"node_id": "a4b15fbe-d158-4265-b5ef-7acf0309dbb3", "node_type": "1", "metadata": {"page_label": "14", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "6d0c1f0fa5ef555549109e642e9a987e4eaf08ec2941df7e7ef9575b54dbc56b"}, "3": {"node_id": "3f8b5798-e6c2-4a6e-b2f0-a8c043396824", "node_type": "1", "metadata": {"page_label": "14", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "2e32b17668687507375d9641ae8888ec8c1c824f6beec7543287b93444894d81"}}, "hash": "2e5bb07417d7a1c0f9a9fec667832144d3ccf5f14121a46cb977eec81ed7066a", "text": "and serves as SPC/PC\nin leading conferences such as IJCAI, AAAI, ICML, NIPS, etc.\nHengShu Zhu (SM\u201919) is currently a principal\ndata scientist &architect at Baidu Inc. He re-\nceived the Ph.D. degree in 2014 and B.E. de-\ngree in 2009, both in Computer Science from\nUniversity of Science and Technology of China\n(USTC), China. His general area of research is\ndata mining and machine learning, with a fo-\ncus on developing advanced data analysis tech-\nniques for innovative business applications. He\nhas published proli\ufb01cally in refereed journals and\nconference proceedings, including IEEE Trans-\nactions on Knowledge and Data Engineering (TKDE), IEEE Transac-\ntions on Mobile Computing (TMC), ACM Transactions on Information\nSystems (ACM TOIS), ACM Transactions on Knowledge Discovery from\nData (TKDD), ACM SIGKDD, ACM SIGIR, WWW, IJCAI, and AAAI. He\nhas served regularly on the organization and program committees of\nnumerous conferences, including as a program co-chair of the KDD\nCup-2019 Regular ML Track, and a founding co-chair of the \ufb01rst In-\nternational Workshop on Organizational Behavior and Talent Analytics\n(OBTA) and the International Workshop on Talent and Management\nComputing (TMC), in conjunction with ACM SIGKDD. He was the re-\ncipient of the Distinguished Dissertation Award of CAS (2016), the\nDistinguished Dissertation Award of CAAI (2016), the Special Prize of\nPresident Scholarship for Postgraduate Students of CAS (2014), the\nBest Student Paper Award of KSEM-2011, WAIM-2013, CCDM-2014,\nand the Best Paper Nomination of ICDM-2014. He is the senior member\nof IEEE, ACM, and CCF .\nXiao-Ru Gao is working towards the Ph.D. de-\ngree in Marketing", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "3f8b5798-e6c2-4a6e-b2f0-a8c043396824": {"__data__": {"id_": "3f8b5798-e6c2-4a6e-b2f0-a8c043396824", "embedding": null, "metadata": {"page_label": "14", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What are the research interests of Xiao-Ru Gao and Hui Xiong?\n2. What are the notable awards and recognitions received by Xiao-Ru Gao and Hui Xiong in their academic and professional careers?\n3. What are the conferences and organizations that Hui Xiong has served on or is currently involved with?", "prev_section_summary": "The section discusses the research focus and contributions of HengShu Zhu, a principal data scientist & architect at Baidu Inc. His research is in the field of data mining and machine learning, with a focus on developing advanced data analysis techniques for innovative business applications. He has published prolifically in refereed journals and conference proceedings, and has served regularly on the organization and program committees of numerous conferences. He has received several awards and recognitions for his research and contributions in the field, including the Distinguished Dissertation Award of CAS and the Best Student Paper Award of several conferences. The section also mentions Xiao-Ru Gao, who is working towards a Ph.D. degree in Marketing.", "section_summary": "The section discusses the research interests, notable awards and recognitions, conferences and organizations served on or involved with, and notable achievements of Xiao-Ru Gao and Hui Xiong. Xiao-Ru Gao's research interests lie primarily in data mining, social networking, influencer marketing, and brand image perception. Hui Xiong's research interests include data mining, social networking, and management practice. Hui Xiong has received numerous awards and recognitions, including the Distinguished Dissertation Award of CAS and CAAI, the Special Prize of President Scholarship for Postgraduate Students of CAS, and the Ram Charan Management Practice Award from the Harvard Business Review. He is also a co-Editor-in-Chief of Encyclopedia of GIS and an Associate Editor of several other journals. He has served regularly on the organization and program committees of numerous conferences, including as a Program Co-Chair of the 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining.", "excerpt_keywords": "Distinguished Dissertation Award, CAS, CAAI, Special Prize, President Scholarship, Best Student Paper, KSEM, WAIM, CCDM, ICDM, IEEE, ACM, CCF, Ph.D., Marketing, Data Mining, Social Networking, Influencer Marketing, Brand Image Perception, Harvard Business Review, RBS Dean's Research Professorship, Rutgers University Board of Trustees Research Fellowship, ICDM Best Research Paper Award, IEEE ICDM Outstanding Service Award, Encyclopedia of GIS, IEEE Transactions on Big Data, ACM Transactions on Knowledge Discovery from Data, ACM Transactions on Management Information Systems, Program Committee, Conference Organization, Program Co-Chair, Industrial and Government Track, 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, IEEE 2013 International Conference on Data Mining, Ph.D. degree, University of Minnesota, USA, Co-Editor-in-Chief, Encyclopedia of GIS, Associate Editor, IEEE Transactions on Big Data, ACM Transactions on Knowledge Discovery"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "b1879836-b599-41e1-af94-4bfe72a77860", "node_type": "4", "metadata": {"page_label": "14", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a8f60bb292a6a8a5a8797722564372a56ffb6de873685c2bc9b524a068f8aafa"}, "2": {"node_id": "856e13fe-3def-4ed6-b440-d474562d6790", "node_type": "1", "metadata": {"page_label": "14", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "2e5bb07417d7a1c0f9a9fec667832144d3ccf5f14121a46cb977eec81ed7066a"}, "3": {"node_id": "8cfb97d7-4310-4249-9ead-f100246bb15f", "node_type": "1", "metadata": {"page_label": "14", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "b900da4d37d0e8dd2c094026cd6322a09a24d509d8719818869308bf7c8b22b5"}}, "hash": "2e32b17668687507375d9641ae8888ec8c1c824f6beec7543287b93444894d81", "text": "He was the re-\ncipient of the Distinguished Dissertation Award of CAS (2016), the\nDistinguished Dissertation Award of CAAI (2016), the Special Prize of\nPresident Scholarship for Postgraduate Students of CAS (2014), the\nBest Student Paper Award of KSEM-2011, WAIM-2013, CCDM-2014,\nand the Best Paper Nomination of ICDM-2014. He is the senior member\nof IEEE, ACM, and CCF .\nXiao-Ru Gao is working towards the Ph.D. de-\ngree in Marketing at Rutgers, the State Univer-\nsity of New Jersey, US. Her research interests lie\nprimarily in data mining, social networking, es-\npecially in\ufb02uencer marketing, and brand image\nperception.\nHui Xiong (Fellow) is currently a Full Profes-\nsor at the Rutgers, the State University of New\nJersey, where he received the 2018 Ram Cha-\nran Management Practice Award as the Grand\nPrix winner from the Harvard Business Review,\nRBS Dean\u2019s Research Professorship (2016), the\nRutgers University Board of Trustees Research\nFellowship for Scholarly Excellence (2009), the\nICDM Best Research Paper Award (2011), and\nthe IEEE ICDM Outstanding Service Award\n(2017). He received the Ph.D. degree from the\nUniversity of Minnesota (UMN), USA. He is a co-Editor-in-Chief of En-\ncyclopedia of GIS, an Associate Editor of IEEE Transactions on Big Data\n(TBD), ACM Transactions on Knowledge Discovery from Data (TKDD),\nand ACM Transactions on Management Information Systems (TMIS).\nHe has served regularly on the organization and program committees\nof numerous conferences, including as a Program Co-Chair of the\nIndustrial and Government Track for the 18th ACM SIGKDD Interna-\ntional Conference on Knowledge Discovery and Data Mining (KDD), a\nProgram Co-Chair for the IEEE 2013 International Conference on Data\nMining (ICDM), a", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "8cfb97d7-4310-4249-9ead-f100246bb15f": {"__data__": {"id_": "8cfb97d7-4310-4249-9ead-f100246bb15f", "embedding": null, "metadata": {"page_label": "14", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is Jian Yang's current academic affiliation and research interests?\n2. What are some of the conferences and organizations that Jian Yang has served on or is currently serving on?\n3. What are some of the scientific papers that Jian Yang has authored and how many times have they been cited in the Web of Science and Scholar Google?", "prev_section_summary": "The section discusses the research interests, notable awards and recognitions, conferences and organizations served on or involved with, and notable achievements of Xiao-Ru Gao and Hui Xiong. Xiao-Ru Gao's research interests lie primarily in data mining, social networking, influencer marketing, and brand image perception. Hui Xiong's research interests include data mining, social networking, and management practice. Hui Xiong has received numerous awards and recognitions, including the Distinguished Dissertation Award of CAS and CAAI, the Special Prize of President Scholarship for Postgraduate Students of CAS, and the Ram Charan Management Practice Award from the Harvard Business Review. He is also a co-Editor-in-Chief of Encyclopedia of GIS and an Associate Editor of several other journals. He has served regularly on the organization and program committees of numerous conferences, including as a Program Co-Chair of the 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining.", "section_summary": "The section provides information about Jian Yang, a researcher in the field of pattern recognition and computer vision. He is currently a Chang-Jiang Professor with the School of Computer Science and Engineering at NUST and has authored more than 200 scientific papers. His papers have been cited more than 6000 times in the Web of Science and 15,000 times in the Scholar Google. He is also an IEEE Fellow and an ACM Distinguished Scientist.\n\nThe section mentions some of the conferences and organizations that Jian Yang has served on or is currently serving on, including the Industrial and Government Track for the 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), the IEEE 2013 International Conference on Data Mining (ICDM), the IEEE 2015 International Conference on Data Mining (ICDM), and the Research Track for the 2018 ACM SIGKDD International Conference on Knowledge Discovery and Data Mining.\n\nThe section also mentions some of the scientific papers that Jian Yang has authored, but does not provide specific details on", "excerpt_keywords": "GIS, Encyclopedia, Associate Editor, IEEE Transactions on Big Data, ACM Transactions on Knowledge Discovery from Data, ACM Transactions on Management Information Systems, Program Co-Chair, ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, IEEE 2013 International Conference on Data Mining, IEEE 2015 International Conference on Data Mining, General Co-Chair, ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, IEEE Fellow, ACM Distinguished Scientist, Pattern Recognition, Computer Vision, Machine Learning, Fellow of IAPR."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "b1879836-b599-41e1-af94-4bfe72a77860", "node_type": "4", "metadata": {"page_label": "14", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a8f60bb292a6a8a5a8797722564372a56ffb6de873685c2bc9b524a068f8aafa"}, "2": {"node_id": "3f8b5798-e6c2-4a6e-b2f0-a8c043396824", "node_type": "1", "metadata": {"page_label": "14", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "2e32b17668687507375d9641ae8888ec8c1c824f6beec7543287b93444894d81"}, "3": {"node_id": "7d0469e6-407b-40f2-8c73-5d72911d8f27", "node_type": "1", "metadata": {"page_label": "14", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a201bceaea49e010ce3e3ada816ccd2f31a79fbd86ce3f9707f9be12b5950a98"}}, "hash": "b900da4d37d0e8dd2c094026cd6322a09a24d509d8719818869308bf7c8b22b5", "text": "of En-\ncyclopedia of GIS, an Associate Editor of IEEE Transactions on Big Data\n(TBD), ACM Transactions on Knowledge Discovery from Data (TKDD),\nand ACM Transactions on Management Information Systems (TMIS).\nHe has served regularly on the organization and program committees\nof numerous conferences, including as a Program Co-Chair of the\nIndustrial and Government Track for the 18th ACM SIGKDD Interna-\ntional Conference on Knowledge Discovery and Data Mining (KDD), a\nProgram Co-Chair for the IEEE 2013 International Conference on Data\nMining (ICDM), a General Co-Chair for the IEEE 2015 International\nConference on Data Mining (ICDM), and a Program Co-Chair of the\nResearch Track for the 2018 ACM SIGKDD International Conference on\nKnowledge Discovery and Data Mining. He is an IEEE Fellow and an\nACM Distinguished Scientist.\nJian Yang (M\u201908) received the Ph.D. degree\nin pattern recognition and intelligence systems\nfrom the Nanjing University of Science and Tech-\nnology (NUST), Nanjing, China, in 2002. In\n2003, he was a Post-Doctoral Researcher with\nthe University of Zaragoza, Zaragoza, Spain.\nFrom 2004 to 2006, he was a Post-Doctoral Fel-\nlow with the Biometrics Centre, The Hong Kong\nPolytechnic University, Hong Kong. From 2006\nto 2007, he was a Post-Doctoral Fellow with the\nDepartment of Computer Science, New Jersey\nInstitute of Technology, Newark, NJ, USA. He is currently a Chang-Jiang\nProfessor with the School of Computer Science and Engineering, NUST.\nHe has authored more than 200 scienti\ufb01c papers in pattern recognition\nand computer vision. His papers have been cited more than 6000 times\nin the Web of Science and 15,000 times in the Scholar Google. His\ncurrent research interests include pattern recognition, computer vision,\nand machine learning. Dr. Y ang is a Fellow of IAPR. He is currently\nan Associate Editor of Pattern Recognition, Pattern Recognition", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "7d0469e6-407b-40f2-8c73-5d72911d8f27": {"__data__": {"id_": "7d0469e6-407b-40f2-8c73-5d72911d8f27", "embedding": null, "metadata": {"page_label": "14", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is Dr. Yi Yang's current research interests and affiliations?\n2. How many scientific papers has Dr. Yi Yang authored and how many times have they been cited in the Web of Science and Scholar Google?\n3. What are the current research interests of Dr. Yi Yang and what journals and conferences does he currently serve as an Associate Editor?", "prev_section_summary": "The section provides information about Jian Yang, a researcher in the field of pattern recognition and computer vision. He is currently a Chang-Jiang Professor with the School of Computer Science and Engineering at NUST and has authored more than 200 scientific papers. His papers have been cited more than 6000 times in the Web of Science and 15,000 times in the Scholar Google. He is also an IEEE Fellow and an ACM Distinguished Scientist.\n\nThe section mentions some of the conferences and organizations that Jian Yang has served on or is currently serving on, including the Industrial and Government Track for the 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), the IEEE 2013 International Conference on Data Mining (ICDM), the IEEE 2015 International Conference on Data Mining (ICDM), and the Research Track for the 2018 ACM SIGKDD International Conference on Knowledge Discovery and Data Mining.\n\nThe section also mentions some of the scientific papers that Jian Yang has authored, but does not provide specific details on", "section_summary": "The section discusses Dr. Yi Yang's current research interests and affiliations, as well as his scientific publications and citations. Dr. Yang is a Chang-Jiang Professor with the School of Computer Science and Engineering at NUST in Newark, NJ, USA. His research interests include pattern recognition, computer vision, and machine learning. He has authored over 200 scientific papers and has been cited more than 6000 times in the Web of Science and 15,000 times in Scholar Google. Dr. Yang is a Fellow of IAPR and serves as an Associate Editor of several journals and conferences, including Pattern Recognition, Pattern Recognition Letters, the IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, and Neurocomputing.", "excerpt_keywords": "pattern recognition, computer vision, machine learning, IAPR, Fellow, Associate Editor, Pattern Recognition, Pattern Recognition Letters, IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, Neurocomputing, New Jersey Institute of Technology, Newark, NJ, USA, Chang-Jiang Professor, School of Computer Science and Engineering, NUST, more than 200 scientific papers, cited more than 6000 times in Web of Science, cited more than 15,000 times in Scholar Google, current research interests."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "b1879836-b599-41e1-af94-4bfe72a77860", "node_type": "4", "metadata": {"page_label": "14", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a8f60bb292a6a8a5a8797722564372a56ffb6de873685c2bc9b524a068f8aafa"}, "2": {"node_id": "8cfb97d7-4310-4249-9ead-f100246bb15f", "node_type": "1", "metadata": {"page_label": "14", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "b900da4d37d0e8dd2c094026cd6322a09a24d509d8719818869308bf7c8b22b5"}}, "hash": "a201bceaea49e010ce3e3ada816ccd2f31a79fbd86ce3f9707f9be12b5950a98", "text": "New Jersey\nInstitute of Technology, Newark, NJ, USA. He is currently a Chang-Jiang\nProfessor with the School of Computer Science and Engineering, NUST.\nHe has authored more than 200 scienti\ufb01c papers in pattern recognition\nand computer vision. His papers have been cited more than 6000 times\nin the Web of Science and 15,000 times in the Scholar Google. His\ncurrent research interests include pattern recognition, computer vision,\nand machine learning. Dr. Y ang is a Fellow of IAPR. He is currently\nan Associate Editor of Pattern Recognition, Pattern Recognition Letters,\nthe IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING\nSYSTEMS, and Neurocomputing.\nAuthorized licensed use limited to: Univ of Calif Santa Barbara. Downloaded on June 20,2021 at 08:36:01 UTC from IEEE Xplore.  Restrictions apply.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "ab357d46-6c15-4fb7-98af-1eec2922bd7f": {"__data__": {"id_": "ab357d46-6c15-4fb7-98af-1eec2922bd7f", "embedding": null, "metadata": {"page_label": "474", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using enterprise architecture (EA) models for decision making in planning purposes?\n2. How does a gap analysis help in identifying areas that need improvement in an enterprise architecture?\n3. What is the role of a transformation model and an action repository in creating transformation paths towards a desired and detailed target architecture?", "prev_section_summary": "The section discusses Dr. Yi Yang's current research interests and affiliations, as well as his scientific publications and citations. Dr. Yang is a Chang-Jiang Professor with the School of Computer Science and Engineering at NUST in Newark, NJ, USA. His research interests include pattern recognition, computer vision, and machine learning. He has authored over 200 scientific papers and has been cited more than 6000 times in the Web of Science and 15,000 times in Scholar Google. Dr. Yang is a Fellow of IAPR and serves as an Associate Editor of several journals and conferences, including Pattern Recognition, Pattern Recognition Letters, the IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, and Neurocomputing.", "section_summary": "The section discusses the use of enterprise architecture (EA) models for decision making in planning purposes. It explains how a gap analysis can help identify areas that need improvement in an enterprise architecture. The role of a transformation model and an action repository in creating transformation paths towards a desired and detailed target architecture is also described. The paper presents a use case example and proposes a technical realization of the solution. Key topics and entities discussed include enterprise architecture management, enterprise architecture, gap analysis, transformation model, action repository, and graph transformation.", "excerpt_keywords": "Enterprise architecture planning, gap analysis, transformation model, graph transformation, Enterprise architecture management, EA models, business, processes, integration, software, technology, managed approach, complexity, abstraction, redesign, enterprise, planning, decision making, change management, adaptation, evolution, organization, components, relationships, environment, principles, guided design, controlled redesign, complexity management."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "e4af41c9-853f-4975-8439-415e32a44762", "node_type": "4", "metadata": {"page_label": "474", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "aee6cf38a4e5eabdb63c856c12d762982faa68814c4c26a216e7e82966f4b677"}, "3": {"node_id": "d268d7b3-399c-43ba-b818-05417e289f3c", "node_type": "1", "metadata": {"page_label": "474", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "b13730e58f2872a15c64afe9f72661c2535fd2df2c368c633b0828a063d67660"}}, "hash": "e088821ba90961cf4c053b97a8eec59e5e7d7ed0145abd7dd917cc2ce1e2240c", "text": "From Gaps to Transformation Paths\nin Enterprise Architecture Planning\nPhilipp Diefenthaler1,2(B)and Bernhard Bauer2\n1Softplant GmbH, Munich, Germany\n2Institute for Software & Systems Engineering, University of Augsburg,\nAugsburg, Germany\nphilipp.diefenthaler@softplant.de,\nBernhard.Bauer@informatik.uni-augsburg.de\nAbstract. Planning changes in an enterprise and its supporting IT\ncan be supported by enterprise architecture (EA) models. The planned\nchanges result in gaps which can be derived by a gap analysis. But, know-ing the gaps is not enough. Also important is to know in which sequence\ngaps are to be closed for transformation path planning. In this paper we\nshow how gaps are identi\ufb01ed and reused for detailing a model of the tar-get architecture. Based on this re\ufb01nement further gaps become visible.\nFurthermore, we describe how it is possible to create with a transfor-\nmation model and an action repository transformation paths towards a\ndesired and detailed target architecture. Afterwards, we give a use case\nexample and propose a technical realization of the solution.\nKeywords: Enterprise architecture planning\n\u00b7Gap analysis \u00b7Transfor-\nmation model \u00b7Graph transformation\n1 Introduction\nEnterprises nowadays face challenges like changing markets, security threats,\nevolving technologies and new regulations that drive the need to adapt theenterprise. Enterprise architecture management (EAM) supports this change\nin a structured manner. An enterprise architecture (EA) is the \u201cfundamental\norganization of a system [the enterprise] embodied in its components, their rela-tionships to each other, and to the environment, and the principles guiding its\ndesign and evolution\u201d [ 1].\nModels of this architecture can support decision making for planning pur-\nposes. Such EA models cover aspects from business, processes, integration, soft-\nware and technology [ 2]. To cope with the complexity of an EA it is crucial\nfor enterprises to use a managed approach to steer and control the redesign ofthe enterprise. The complexity arises from the level of abstraction, the", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "d268d7b3-399c-43ba-b818-05417e289f3c": {"__data__": {"id_": "d268d7b3-399c-43ba-b818-05417e289f3c", "embedding": null, "metadata": {"page_label": "474", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What are the key components of an Enterprise Architecture (EA) model and how do they support decision making for planning purposes?\n2. How does the complexity of an EA arise and what challenges does it present for enterprises in terms of steering and controlling its redesign?\n3. What is the role of a managed approach in EA planning and how can it help enterprises cope with the complexity of their systems?", "prev_section_summary": "The section discusses the use of enterprise architecture (EA) models for decision making in planning purposes. It explains how a gap analysis can help identify areas that need improvement in an enterprise architecture. The role of a transformation model and an action repository in creating transformation paths towards a desired and detailed target architecture is also described. The paper presents a use case example and proposes a technical realization of the solution. Key topics and entities discussed include enterprise architecture management, enterprise architecture, gap analysis, transformation model, action repository, and graph transformation.", "section_summary": "The section discusses the key components of an Enterprise Architecture (EA) model and their role in supporting decision making for planning purposes. It also explores the complexity of an EA and the challenges it presents for enterprises in terms of steering and controlling its redesign. The section highlights the importance of a managed approach in EA planning and how it can help enterprises cope with the complexity of their systems. The section also discusses the level of abstraction, the number of stakeholders involved, and the change of internal and external conditions inherent to EAs as factors contributing to the complexity of an EA.", "excerpt_keywords": "Enterprise architecture, decision making, planning, complexity, managed approach, stakeholders, abstraction, change, internal conditions, external conditions, software, technology."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "e4af41c9-853f-4975-8439-415e32a44762", "node_type": "4", "metadata": {"page_label": "474", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "aee6cf38a4e5eabdb63c856c12d762982faa68814c4c26a216e7e82966f4b677"}, "2": {"node_id": "ab357d46-6c15-4fb7-98af-1eec2922bd7f", "node_type": "1", "metadata": {"page_label": "474", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "e088821ba90961cf4c053b97a8eec59e5e7d7ed0145abd7dd917cc2ce1e2240c"}}, "hash": "b13730e58f2872a15c64afe9f72661c2535fd2df2c368c633b0828a063d67660", "text": "is the \u201cfundamental\norganization of a system [the enterprise] embodied in its components, their rela-tionships to each other, and to the environment, and the principles guiding its\ndesign and evolution\u201d [ 1].\nModels of this architecture can support decision making for planning pur-\nposes. Such EA models cover aspects from business, processes, integration, soft-\nware and technology [ 2]. To cope with the complexity of an EA it is crucial\nfor enterprises to use a managed approach to steer and control the redesign ofthe enterprise. The complexity arises from the level of abstraction, the num-\nber of stakeholders involved, and the change of internal and external conditions\ninherent to EAs.\nc\u20ddSpringer International Publishing Switzerland 2014\nS. Hammoudi et al. (Eds.): ICEIS 2013, LNBIP 190, pp. 474\u2013489, 2014.DOI: 10.1007/978-3-319-09492-2\n28", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "81acb7c7-dcff-4da1-b3d0-259f2d86fa79": {"__data__": {"id_": "81acb7c7-dcff-4da1-b3d0-259f2d86fa79", "embedding": null, "metadata": {"page_label": "475", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What are the different decision levels involved in EA planning activities and how do they vary in detail and levels of abstraction?\n2. How can gaps be derived from two EA models for different points in time, and what is the transformation model by Aier and Gleichauf used to connect architectural building blocks from these models?\n3. What is the proposed technical realization for the solution to get from gaps to transformation paths based on semantic web technologies and graph transformations?", "prev_section_summary": "The section discusses the key components of an Enterprise Architecture (EA) model and their role in supporting decision making for planning purposes. It also explores the complexity of an EA and the challenges it presents for enterprises in terms of steering and controlling its redesign. The section highlights the importance of a managed approach in EA planning and how it can help enterprises cope with the complexity of their systems. The section also discusses the level of abstraction, the number of stakeholders involved, and the change of internal and external conditions inherent to EAs as factors contributing to the complexity of an EA.", "section_summary": "The section discusses the different decision levels involved in Enterprise Architecture (EA) planning activities and how they vary in detail and levels of abstraction. It also describes how gaps can be derived from two EA models for different points in time and introduces a transformation model by Aier and Gleichauf to connect architectural building blocks from these models. The section proposes a technical realization based on semantic web technologies and graph transformations to get from gaps to transformation paths. The foundations of EA models and their usage for planning purposes are also introduced, along with semantic web technologies and graph transformations for planning purposes.", "excerpt_keywords": "Enterprise Architecture, Planning, Models, Gaps, Transformation, Architectural Building Blocks, Action Repository, Semantic Web Technologies, Graph Transformations, Management Cycle."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "3c2e8b58-e9eb-4f7a-884f-41f0155d56ef", "node_type": "4", "metadata": {"page_label": "475", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "2f7788aa4b917d49a21e64d362ffc8d6188998d57f3ba44afd05a23057110c4d"}, "3": {"node_id": "fc8366b8-f6a1-4bd1-ade4-041f326ac4dc", "node_type": "1", "metadata": {"page_label": "475", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "7172efb129356757dd5c7aaa5e79680c5b63afa51ea9cdfb5fd5204006bb38f9"}}, "hash": "b9c6dafee08311273588d181f3bb8b57af6154afbb095703e99cdb0e01a5564b", "text": "From Gaps to Transformation Paths in Enterprise Architecture Planning 475\nTo plan the change it is necessary to have a plan basis, i.e. the current\narchitecture, and to know the goal of planning activities, i.e. the target architec-ture. According to [ 3,4] the planning activities take place at di\ufb00erent decision\nlevels. Each of them varies in detail and levels of abstraction seem to be inevitable\n[4]. The need to change and the resulting moving target are challenges for EA\nplanning, as part of the EAM, has to meet [ 5,6]. EAM and particularly EA plan-\nning is supported by tools which allow the creation of visualizations, automated\ndocumentation and analysis of EA models.\nIn this paper we describe how gaps can be derived from two EA models for\ndi\ufb00erent points in time. Furthermore, we introduce the transformation model by\nAier and Gleichauf [ 7] to connect architectural building blocks from the models\nof the current and target EA. With the results from gap analysis and the infor-\nmation contained in the transformation model we introduce an action repository\nfor the creation of di\ufb00erent transformation paths. We exemplify the solution to\nget from gaps to transformation paths based on a model of a current and target\narchitecture of an application architecture within a use case for a master dataconsolidation challenge. Furthermore, we propose a technical realization based\non semantic web technologies and graph transformations.\n2 Foundations\nThis section gives an introduction to the foundations of EA models and their\nusage for planning purposes. Furthermore, we introduce semantic web technolo-\ngies and graph transformations for planning purposes, as they are of relevance\nfor our proposed technical realization of the solution.\n2.1 Enterprise Architecture Models\nAccording to Buckl and Schweda [ 8] EAM follows a typical management cycle\nthat consists of the phases plan, do, check and act. The plan phase is concerned\nwith developing change proposals that are implemented in the do phase. Within\nthe check phase di\ufb00erences between intended and actually achieved results arecontrolled. Based upon the results from the check phase the act phase", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "fc8366b8-f6a1-4bd1-ade4-041f326ac4dc": {"__data__": {"id_": "fc8366b8-f6a1-4bd1-ade4-041f326ac4dc", "embedding": null, "metadata": {"page_label": "475", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the typical management cycle for Enterprise Architecture Models (EAM) and how do models support the plan phase of this cycle?\n2. How can EA models be used to describe an enterprise's architecture at different points in time and guide the development of an EA from the current towards a target architecture?\n3. What are some common EA goals that influence the development of a target architecture and how do they impact the process of master data consolidation, IT flexibility, and standard platform coverage?", "prev_section_summary": "The section discusses the different decision levels involved in Enterprise Architecture (EA) planning activities and how they vary in detail and levels of abstraction. It also describes how gaps can be derived from two EA models for different points in time and introduces a transformation model by Aier and Gleichauf to connect architectural building blocks from these models. The section proposes a technical realization based on semantic web technologies and graph transformations to get from gaps to transformation paths. The foundations of EA models and their usage for planning purposes are also introduced, along with semantic web technologies and graph transformations for planning purposes.", "section_summary": "The section discusses the management cycle for Enterprise Architecture Models (EAM) and how models can support the plan phase of this cycle. EA models can be used to describe an enterprise's architecture at different points in time and guide the development of an EA from the current towards a target architecture. The development of a target architecture is influenced by business requirements, strategic goals, and IT objectives such as master data consolidation, improving IT flexibility, and driving the coverage of standard platforms. The section also introduces semantic web technologies and graph transformations for planning purposes.", "excerpt_keywords": "Enterprise Architecture Models, Management Cycle, Change Proposals, Implementation, Check Phase, Act Phase, Abstraction Mechanism, EA Goals, Business Requirements, Strategic Goals, IT Objectives, Master Data Consolidation, Flexibility, Standard Platforms."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "3c2e8b58-e9eb-4f7a-884f-41f0155d56ef", "node_type": "4", "metadata": {"page_label": "475", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "2f7788aa4b917d49a21e64d362ffc8d6188998d57f3ba44afd05a23057110c4d"}, "2": {"node_id": "81acb7c7-dcff-4da1-b3d0-259f2d86fa79", "node_type": "1", "metadata": {"page_label": "475", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "b9c6dafee08311273588d181f3bb8b57af6154afbb095703e99cdb0e01a5564b"}}, "hash": "7172efb129356757dd5c7aaa5e79680c5b63afa51ea9cdfb5fd5204006bb38f9", "text": "their\nusage for planning purposes. Furthermore, we introduce semantic web technolo-\ngies and graph transformations for planning purposes, as they are of relevance\nfor our proposed technical realization of the solution.\n2.1 Enterprise Architecture Models\nAccording to Buckl and Schweda [ 8] EAM follows a typical management cycle\nthat consists of the phases plan, do, check and act. The plan phase is concerned\nwith developing change proposals that are implemented in the do phase. Within\nthe check phase di\ufb00erences between intended and actually achieved results arecontrolled. Based upon the results from the check phase the act phase provides\ninput to the plan phase by supplying information for the next plan phase. Models,\nas an abstraction mechanism, of an enterprise, can support the plan phase as\npart of an EAM approach [ 9,10].\nEA models can be used to describe an EA for di\ufb00erent points in time [ 8]. The\nmodel of the current architecture of the enterprise is a documented architecture\nat the present point in time and serves as a starting point for de\ufb01ning a model of\na target architecture. In contrast the model of the target architecture representsa desired architecture in the future which can be used to guide the development\nof an EA from the current towards a target architecture. The development of\na target architecture depends on the enterprises\u2019 EA goals. It is in\ufb02uenced bybusiness requirements, strategic goals and IT objectives like master data con-\nsolidation, improving the \ufb02exibility of IT and drive the coverage of standard\nplatforms [ 11].", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "813b7a2f-e1b4-4784-96f3-1be0370c1b22": {"__data__": {"id_": "813b7a2f-e1b4-4784-96f3-1be0370c1b22", "embedding": null, "metadata": {"page_label": "476", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the difference between a gap analysis and a delta analysis in enterprise architecture planning?\n2. How do semantic web technologies help integrate heterogeneous data sets and formalize the underlying structure of information?\n3. What are the two standards relevant for a proposed technical realization using semantic web technologies?", "prev_section_summary": "The section discusses the management cycle for Enterprise Architecture Models (EAM) and how models can support the plan phase of this cycle. EA models can be used to describe an enterprise's architecture at different points in time and guide the development of an EA from the current towards a target architecture. The development of a target architecture is influenced by business requirements, strategic goals, and IT objectives such as master data consolidation, improving IT flexibility, and driving the coverage of standard platforms. The section also introduces semantic web technologies and graph transformations for planning purposes.", "section_summary": "The section discusses the difference between a gap analysis and a delta analysis in enterprise architecture planning, and how semantic web technologies help integrate heterogeneous data sets and formalize the underlying structure of information. The two standards relevant for a proposed technical realization using semantic web technologies are the Web Ontology Language (OWL) and the SPARQL Query Language for RDF. The Resource Description Framework (RDF) is a basis for both standards, and an RDF graph consists of triples of the form 'subject, predicate, object'.", "excerpt_keywords": "1. Architecture method, 2. Gap analysis, 3. Semantic web technologies, 4. Ontology, 5. Web Ontology Language (OWL), 6. SPARQL Query Language, 7. Resource Description Framework (RDF), 8. Information querying, 9. Machine understanding, 10. Heterogeneous data sets."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "9d4c0667-0535-499e-afcc-b195f78cb549", "node_type": "4", "metadata": {"page_label": "476", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "5de25c19315462d3549f5e784b49d5fe540f2dd41be784136ea8f183d044841e"}, "3": {"node_id": "f00a11b4-bbda-48c0-a484-294f58a4e657", "node_type": "1", "metadata": {"page_label": "476", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "5c4540d2a91fa3dc4ef4ce92a346d2f51d39e554ebc72f081e6d4be6a6e5a440"}}, "hash": "29e6f47010061b9a69ff8ddf02fee954e777e87990df9d7eac17ea5f987f34f3", "text": "476 P. Diefenthaler and B. Bauer\nWhich factors and how exactly they in\ufb02uence the target architecture depends\non the architecture method applied and how it is integrated into the enterprise\u2018sgovernance processes.\nA gap analysis, sometimes also referred to as delta analysis, is the comparison\nbetween two models of an EA that is used to clarify the di\ufb00erences betweenthose two architectures. Di\ufb00erent models of architectures that can be compared\nare current to target, current to planned, planned to target and planned to\nplanned [ 8].\n2.2 Semantic Web Technologies in a Nutshell\nSemantic web technologies are used to integrate heterogeneous data sets and\nformalize the underlying structure of the information to allow a machine to\nunderstand the semantics of it [ 12]. The World Wide Web Consortium (W3C)\nprovides a set of standards to describe an ontology and to query it. An ontology\u201cis a set of precise descriptive statements about some part of the world (usually\nreferred to as the domain of interest or the subject matter of the ontology)\u201d [ 13].\nTwo standards are of relevance for a proposed technical realization: \ufb01rstly,\nthe Web Ontology Language (OWL) [ 13] for making descriptive statements and\nsecondly, the SPARQL Query Language for RDF (SPARQL) [ 14], which allows\nquerying these statements.\nThe Resource Description Framework (RDF) [ 15] is a basis for both stan-\ndards, as OWL ontologies can be serialized as RDF graphs and can be accessed\nvia SPARQL. An RDF graph consists of triples of the form \u2018subject, predicate,object\u2019, whereas subjects and objects are nodes and predicates are relations.\nEvery resource in an ontology is identi\ufb01ed by a resource identi\ufb01er which allows\nfor example distinguishing between a bank in a \ufb01nancial context and a bank ofa river. Information from the ontology is queried via SPARQL, which provides\nthe resources that match patterns speci\ufb01ed within", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "f00a11b4-bbda-48c0-a484-294f58a4e657": {"__data__": {"id_": "f00a11b4-bbda-48c0-a484-294f58a4e657", "embedding": null, "metadata": {"page_label": "476", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the difference between an RDF graph and a task network in planning approaches for Enterprise Architecture?\n2. How have semantic web technologies been applied to domains of interest, including EAM, and what are some existing implementations?\n3. What are some state space model based planning approaches used in EAM and how are they different from task networks?", "prev_section_summary": "The section discusses the difference between a gap analysis and a delta analysis in enterprise architecture planning, and how semantic web technologies help integrate heterogeneous data sets and formalize the underlying structure of information. The two standards relevant for a proposed technical realization using semantic web technologies are the Web Ontology Language (OWL) and the SPARQL Query Language for RDF. The Resource Description Framework (RDF) is a basis for both standards, and an RDF graph consists of triples of the form 'subject, predicate, object'.", "section_summary": "The section discusses the use of graph neural networks in planning approaches for Enterprise Architecture (EAM). It explains the difference between an RDF graph and a task network in planning approaches for EAM, and how semantic web technologies have been applied to domains of interest, including EAM. The section also discusses state space model based planning approaches used in EAM and how they differ from task networks. Finally, the section mentions some existing implementations of semantic web technologies for EAM, including TopQuadrant's TopBraid Composer and Essential Project.", "excerpt_keywords": "Semantic web, ontologies, RDF, SPARQL, planning, state space model, task networks, architecture, EAM, TopBraid Composer, Essential Project, semantic business process modeling, diagnosis of embedded systems, TopQuadrant, state space based approach, models, current architecture, target architecture, tools used in practice, whitepaper, building semantic EAM solutions."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "9d4c0667-0535-499e-afcc-b195f78cb549", "node_type": "4", "metadata": {"page_label": "476", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "5de25c19315462d3549f5e784b49d5fe540f2dd41be784136ea8f183d044841e"}, "2": {"node_id": "813b7a2f-e1b4-4784-96f3-1be0370c1b22", "node_type": "1", "metadata": {"page_label": "476", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "29e6f47010061b9a69ff8ddf02fee954e777e87990df9d7eac17ea5f987f34f3"}}, "hash": "5c4540d2a91fa3dc4ef4ce92a346d2f51d39e554ebc72f081e6d4be6a6e5a440", "text": "ontologies can be serialized as RDF graphs and can be accessed\nvia SPARQL. An RDF graph consists of triples of the form \u2018subject, predicate,object\u2019, whereas subjects and objects are nodes and predicates are relations.\nEvery resource in an ontology is identi\ufb01ed by a resource identi\ufb01er which allows\nfor example distinguishing between a bank in a \ufb01nancial context and a bank ofa river. Information from the ontology is queried via SPARQL, which provides\nthe resources that match patterns speci\ufb01ed within the query.\nSemantic web technologies have already been applied to domains of interest\nthat range from semantic business process modeling [ 16] to diagnosis of embed-\nded systems [ 17]. First implementations based upon semantic web technologies\nfor EAM already exist from TopQuadrant with its TopBraid Composer\n1and\nEssential Project2.\n2.3 Graph Transformations for Planning Purposes\nSeveral di\ufb00erent approaches, techniques and representations to planning prob-\nlems have been developed over the last decades [ 18,19]. These approaches range\nfrom state space model based planning to task networks, where tasks for reachinga goal are decomposed and sequenced. A state space based approach is prefer-\nable, because models of the current and target architecture are used in many\nEAM approaches [ 5,11,20,21] and are present in tools used in practice [ 22].\n1www.topquadrant.com/docs/whitepapers/WP-BuildingSemanticEASolutions\n-withTopBraid.pdf\n2www.enterprise-architecture.org/", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "3b03f2dd-94df-43a3-9f25-751fd859ee56": {"__data__": {"id_": "3b03f2dd-94df-43a3-9f25-751fd859ee56", "embedding": null, "metadata": {"page_label": "477", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using graph transformations in enterprise architecture planning?\n2. How does the proposed approach of modeling current and target architecture help in generating a transformation path?\n3. What is the role of business support maps in ensuring the comparability of the models in the proposed approach?", "prev_section_summary": "The section discusses the use of graph neural networks in planning approaches for Enterprise Architecture (EAM). It explains the difference between an RDF graph and a task network in planning approaches for EAM, and how semantic web technologies have been applied to domains of interest, including EAM. The section also discusses state space model based planning approaches used in EAM and how they differ from task networks. Finally, the section mentions some existing implementations of semantic web technologies for EAM, including TopQuadrant's TopBraid Composer and Essential Project.", "section_summary": "The section discusses the use of graph transformations in enterprise architecture planning. Graph transformations are used to solve planning problems by applying graph transformations to a model until a solution is found. The proposed approach involves modeling the current and target architecture at the same level of detail to ensure comparability. Business support maps are used to relate applications to supported processes and organization units. The result of the modeling phase are two sets: currentArchitecture and targetArchitecture.", "excerpt_keywords": "1. Graph transformations\n2. Planning problem\n3. State space\n4. Computation time\n5. Graph patterns\n6. Target architecture\n7. Domain expert\n8. Modeling\n9. Business support maps\n10. EA model"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "fc130a71-3273-495a-8425-cb428aafe536", "node_type": "4", "metadata": {"page_label": "477", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "d167f69365feeebf08d4e5311c0293ff15b57098d968240bfde1be734a32b9a1"}, "3": {"node_id": "560eaa93-8cb6-4f32-b481-beb417607cf5", "node_type": "1", "metadata": {"page_label": "477", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "de9959748126d2a010c2c18db36050f1c968cca1fe836d281300b8976908b140"}}, "hash": "d7838bf6e499d86183cc627da2807bedc070a3877130d3a20d5217769b1f4d29", "text": "From Gaps to Transformation Paths in Enterprise Architecture Planning 477\nGraph transformations for AI planning purposes solve a planning problem\nby applying graph transformations to a model until a solution for the planningproblem is found. The result of such a planning process is a sequence of actions\nchanging a model into another model.\nHowever, graph transformations have the disadvantage that they provide a\nhuge state space regarding the states, which have to be examined when all states\nin the graph are computed. As a consequence this in\ufb02uences the computation\ntime of all possible worlds created through the transformations. With graphtransformations a planning problem can be solved by searching for graph pat-\nterns in a state represented by a graph and applying graph transformations to\nchange the state [ 23]. Graph transformations have the bene\ufb01t that they have a\nsound theoretical foundation [ 24].\n3 From Gaps to Transformation Paths as Sequences\nof Actions\nThe goal of the proposed approach is to deliver a more detailed model of the\ntarget architecture by making suggestions to a domain expert how a detailed\ntarget architecture could look like. Afterwards, we describe how these gaps are\nrelated to each other to generate a transformation path which allows to structurechange activities, which close gaps, in sequence of actions.\n3.1 Modeling Current and Target Architecture\nFirst, a current architecture is modeled and afterwards, a target architecture is\nmodeled, at the same level of detail. We reuse the model of the current architec-\nture and change it to the desired target architecture. The same level of detail is\nnecessary to ensure the comparability of the models.\nThe current architecture may be more detailed, but can be aggregated in a\nway which restores the comparability [ 25]. Business support maps, which relate\napplications to supported processes and organization units, are an example forsuch a model with the same level of detail [ 11].\nResults of the Modeling. The result of this phase are the two sets:\ncurrentArchitecture =model of the current architecture of the EA\ntargetArchitecture =model of the target architecture of the EA\nIn our solution the core of an EA model is a set which consists of three", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "560eaa93-8cb6-4f32-b481-beb417607cf5": {"__data__": {"id_": "560eaa93-8cb6-4f32-b481-beb417607cf5", "embedding": null, "metadata": {"page_label": "477", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of business support maps in restoring the comparability of models in enterprise architecture planning?\n2. How does the core of an enterprise architecture model consist of different types of elements, and what are the three types of elements?\n3. What is the difference between the current architecture and target architecture in enterprise architecture planning, and how are they modeled in the solution?", "prev_section_summary": "The section discusses the use of graph transformations in enterprise architecture planning. Graph transformations are used to solve planning problems by applying graph transformations to a model until a solution is found. The proposed approach involves modeling the current and target architecture at the same level of detail to ensure comparability. Business support maps are used to relate applications to supported processes and organization units. The result of the modeling phase are two sets: currentArchitecture and targetArchitecture.", "section_summary": "The section discusses the purpose of business support maps in restoring the comparability of models in enterprise architecture planning. It also explains the core of an enterprise architecture model consisting of three types of elements: architecture building blocks, relations between architecture building blocks, and attributes of architecture building blocks. The section also explains the difference between the current architecture and target architecture in enterprise architecture planning and how they are modeled in the solution.", "excerpt_keywords": "EA, architecture, building blocks, relations, attributes, model, core, sets, elements, comparability, business support maps."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "fc130a71-3273-495a-8425-cb428aafe536", "node_type": "4", "metadata": {"page_label": "477", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "d167f69365feeebf08d4e5311c0293ff15b57098d968240bfde1be734a32b9a1"}, "2": {"node_id": "3b03f2dd-94df-43a3-9f25-751fd859ee56", "node_type": "1", "metadata": {"page_label": "477", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "d7838bf6e499d86183cc627da2807bedc070a3877130d3a20d5217769b1f4d29"}}, "hash": "de9959748126d2a010c2c18db36050f1c968cca1fe836d281300b8976908b140", "text": "the comparability of the models.\nThe current architecture may be more detailed, but can be aggregated in a\nway which restores the comparability [ 25]. Business support maps, which relate\napplications to supported processes and organization units, are an example forsuch a model with the same level of detail [ 11].\nResults of the Modeling. The result of this phase are the two sets:\ncurrentArchitecture =model of the current architecture of the EA\ntargetArchitecture =model of the target architecture of the EA\nIn our solution the core of an EA model is a set which consists of three di\ufb00erent\ntypes of elements. The EA model contains the architecture building blocks (B)\nof the EA, relations between architecture building blocks (R)and attributes of\narchitecture building blocks (A). In this sense an EA model can be de\ufb01ned as:\nM:={B\u222aR\u222aA}\nB:={x|xis an architecture building block }\nR:={x|x\u2208B\u00d7B}and A:={x|x\u2208B\u00d7V}", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "48260b1b-0415-4f80-b0e2-30144e171771": {"__data__": {"id_": "48260b1b-0415-4f80-b0e2-30144e171771", "embedding": null, "metadata": {"page_label": "478", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of performing a gap analysis in architecture building blocks?\n2. How does the successor relationship model work in a transformation model for architecture building blocks?\n3. What are the three subsets identified during a gap analysis in architecture building blocks?", "prev_section_summary": "The section discusses the purpose of business support maps in restoring the comparability of models in enterprise architecture planning. It also explains the core of an enterprise architecture model consisting of three types of elements: architecture building blocks, relations between architecture building blocks, and attributes of architecture building blocks. The section also explains the difference between the current architecture and target architecture in enterprise architecture planning and how they are modeled in the solution.", "section_summary": "The section discusses the purpose of performing a gap analysis in architecture building blocks, how the successor relationship model works in a transformation model for architecture building blocks, and the three subsets identified during a gap analysis in architecture building blocks. The section also explains the definition of the transformation model and the successor relationships.", "excerpt_keywords": "architecture, building blocks, relations, attributes, gap analysis, current architecture, target architecture, stable, transformation model, successor relationships."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "9bc994d8-83ee-4424-aa7d-cab1758f2d15", "node_type": "4", "metadata": {"page_label": "478", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "82b05d0d1fd143fc24f18c451af7e1ea6dd510b3940358ac556bdeba9575a469"}, "3": {"node_id": "f8d21845-0325-4ffa-a0b5-0718e6079c93", "node_type": "1", "metadata": {"page_label": "478", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "3a9ab2018d9efdb8823c3769fe77a9a66946a93a3bc0c8dd835e30647e09baf4"}}, "hash": "5aa51cff56be57baa4cb10c1c1cdacf09b63b5bff9d356db9227b35e33e353ba", "text": "478 P. Diefenthaler and B. Bauer\nArchitecture building blocks stand for the elements of the EA, for instance\na Customer Relationship Management application within the application archi-tecture. Relations hold between these architecture building blocks, for exam-\nple when an application depends on another application the respective building\nblocks are connected by a dependency relation. Attributes are values associatedwith architecture building blocks that characterize measurable and observable\ncharacteristics of the architecture building block, e.g. the release number of an\napplication or the uptime of an service.\n3.2 Performing Gap Analysis\nGap analysis is performed to compare the modeled current and target architec-\nture. It compares the di\ufb00erences between currentArchitecture and targetArchi-\ntecture . In terms of a set operation this comparison corresponds to a intersection\nof the two compared sets. As a result three subsets are identi\ufb01ed: onlyCurrentAr-\nchitecture ,onlyTargetArchitecture and stable.\nResults of Gap Analysis. onlyCurrentArchitecture is the set of building\nblocks, relations and attributes which only exist in the model of the current\narchitecture.\nonlyCurrentArchitecture :={x|x\u2208currentArchitecture\n\u2227x/\u2208targetArchitecture }\nIn contrast, onlyTargetArchitecture is the set of building blocks, relations and\nattributes which only exist in the target architecture.\nonlyT argetArchitecture :={x|x/\u2208currentArchitecture\n\u2227x\u2208targetArchitecture }\nThe third set stable is the set of building blocks, relations and attributes which\nthe current and target architecture have in common.\nstable :={x|x\u2208currentArchitecture \u2227x\u2208targetArchitecture }\n3.3 Setting the Successor Relationships for Building Blocks\nThe successor relationships are modelled within a transformation model [ 7]. The\ntransformation model is de\ufb01ned as follows: transf ormationM odel :={x|x\u2208\ncurrentArchitecture \u00d7targetArchitecture }\nWith the successor relationships at hand it", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "f8d21845-0325-4ffa-a0b5-0718e6079c93": {"__data__": {"id_": "f8d21845-0325-4ffa-a0b5-0718e6079c93", "embedding": null, "metadata": {"page_label": "478", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the definition of the transformation model in the context of building block successor relationships in enterprise architecture planning?\n2. How can the successor relationships be used to identify the successor type for building blocks in the current and target architectures?\n3. What are the different types of successor relationships that can be identified in the transformation model, and how are they defined?", "prev_section_summary": "The section discusses the purpose of performing a gap analysis in architecture building blocks, how the successor relationship model works in a transformation model for architecture building blocks, and the three subsets identified during a gap analysis in architecture building blocks. The section also explains the definition of the transformation model and the successor relationships.", "section_summary": "The section discusses the transformation model in the context of building block successor relationships in enterprise architecture planning. The transformation model is defined as the set of building blocks, relations, and attributes that are common to the current and target architectures. Successor relationships are used to identify the successor type for building blocks in the current and target architectures. The different types of successor relationships that can be identified in the transformation model are noSuccessor, noPredecessor, oneToOne, oneToMany, manyToOne, and manyToMany. The inverse of the successor relation is the predecessor relation. The section also mentions the importance of identifying the successor type for building blocks in order to plan for the transformation of the architecture.", "excerpt_keywords": "1. Building blocks,\n2. Relations,\n3. Attributes,\n4. Transformation model,\n5. Successor relationships,\n6. NoSuccessor,\n7. NoPredecessor,\n8. OneToOne,\n9. OneToMany,\n10. ManyToOne,\n11. ManyToMany."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "9bc994d8-83ee-4424-aa7d-cab1758f2d15", "node_type": "4", "metadata": {"page_label": "478", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "82b05d0d1fd143fc24f18c451af7e1ea6dd510b3940358ac556bdeba9575a469"}, "2": {"node_id": "48260b1b-0415-4f80-b0e2-30144e171771", "node_type": "1", "metadata": {"page_label": "478", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "5aa51cff56be57baa4cb10c1c1cdacf09b63b5bff9d356db9227b35e33e353ba"}}, "hash": "3a9ab2018d9efdb8823c3769fe77a9a66946a93a3bc0c8dd835e30647e09baf4", "text": "}\nThe third set stable is the set of building blocks, relations and attributes which\nthe current and target architecture have in common.\nstable :={x|x\u2208currentArchitecture \u2227x\u2208targetArchitecture }\n3.3 Setting the Successor Relationships for Building Blocks\nThe successor relationships are modelled within a transformation model [ 7]. The\ntransformation model is de\ufb01ned as follows: transf ormationM odel :={x|x\u2208\ncurrentArchitecture \u00d7targetArchitecture }\nWith the successor relationships at hand it is possible to identify the succes-\nsor type for building blocks which can be divided into noSuccessor ,noPredeces-\nsor,oneToOne ,oneToMany ,\nmanyToOne ,a n d manyToMany . The inverse of the\nsuccessor relation is the predecessor relation.\nAll building blocks in onlyCurrentArchitecture that do not have a successor\nb e l o n gt ot h es e t noSuccessor . All building blocks that belong to onlyTargetAr-\nchitecture and do not have a predecessor belong to the set noPredecessor . The set\noneToOne consists of the pairs of building blocks that have exactly one successor\nand this successor has only one predecessor.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "7e4e6d40-6aa8-4450-845d-c01cb3d0879c": {"__data__": {"id_": "7e4e6d40-6aa8-4450-845d-c01cb3d0879c", "embedding": null, "metadata": {"page_label": "479", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the noSuccessor and noPredecessor sets in the transformation model of the current architecture?\n2. How does the model of the current architecture consider applications, services, and business building blocks to generate suggestions for a detailed target architecture?\n3. What is the process for overruling inappropriate suggestions made by the model of the detailed target architecture for a domain expert?", "prev_section_summary": "The section discusses the transformation model in the context of building block successor relationships in enterprise architecture planning. The transformation model is defined as the set of building blocks, relations, and attributes that are common to the current and target architectures. Successor relationships are used to identify the successor type for building blocks in the current and target architectures. The different types of successor relationships that can be identified in the transformation model are noSuccessor, noPredecessor, oneToOne, oneToMany, manyToOne, and manyToMany. The inverse of the successor relation is the predecessor relation. The section also mentions the importance of identifying the successor type for building blocks in order to plan for the transformation of the architecture.", "section_summary": "The section discusses the purpose of the noSuccessor and noPredecessor sets in the transformation model of a current architecture, how the model considers applications, services, and business building blocks to generate suggestions for a detailed target architecture, and the process for overruling inappropriate suggestions made by the model for a domain expert. The section also explains the different subsets of building blocks in the transformation model and how the model generates suggestions for a detailed target architecture by following a suggestion and processing all sets of successor relationships.", "excerpt_keywords": "1. noSuccessor set\n2. provided services\n3. target architecture\n4. applications\n5. dependencies\n6. detailed target architecture\n7. successor relationships\n8. business building blocks\n9. implementation independent\n10. domain expert"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "bd8e03bc-e601-4841-824f-ed9aeec49953", "node_type": "4", "metadata": {"page_label": "479", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "3f51f64cdd0021155737e7826c409d607939518f65e164e3ee555f38594170fc"}, "3": {"node_id": "8ee342f5-c56f-40b6-8ab3-374f48f95ecc", "node_type": "1", "metadata": {"page_label": "479", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "217e6b4b19a89b7c20815788fc428cd116acfcf20bca540432079de9f4979f13"}}, "hash": "087f635b21ff7594f883b7e019bad64e1ca1d37d10959f306b65e18768dc9537", "text": "From Gaps to Transformation Paths in Enterprise Architecture Planning 479\noneToMany is the set of building blocks that have several successors in the\ntarget architecture whereas the set manyToOne is the set of building blocks\nwhich have the same successor in the target architecture. manyToMany is the\nset of building blocks which have common successors, which in turn have several\npredecessors. By querying the models we can determine to which set a buildingblock belongs.\nA successor relationship is part of exactly one of the above subsets. Within\nthe six di\ufb00erent sets disjoint subsets exist. For the noSuccessor and noPredeces-\nsorset each building block represents a disjoint subset and are planned inde-\npendently in contrast to the other successor sets. This is an implicit information\nof the transformation model, as we do not model self-directed relations for thisinformation.\n3.4 Creating Suggestions for a Detailed Target Architecture\nIn order to make suggestions the model of the current architecture considers\napplications, services and business building blocks. Business Building Blocks are\nin a tight relationship with the business activities of an enterprise but implemen-tation independent. With the detailed information of the current architecture\nand the successor relationships at hand for applications it is possible to generate\nsuggestions how a model of a detailed target architecture could look like.\nEach application belongs to exactly one subset of the transformation model.\nDi\ufb00erent suggestions are made for the subsets how to detail the target archi-\ntecture. By following a suggestion the target is stepwise getting more detailed,as all sets of successor relationships are getting processed. A suggestion may\nbe inappropriate for a domain expert she can overrule it by modeling di\ufb00erent\ndetails. The result is a model of a detailed target architecture. At \ufb01rst all ser-vices are transferred to the model of a detailed target architecture. Then the\ndependencies can be added to the model of the detailed target architecture.\nSuggestions for Provided Services.\n1.noSuccessor set: for each provided service in the current architecture check\nif it is used by an application that is part of the target", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "8ee342f5-c56f-40b6-8ab3-374f48f95ecc": {"__data__": {"id_": "8ee342f5-c56f-40b6-8ab3-374f48f95ecc", "embedding": null, "metadata": {"page_label": "479", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation?\n2. How does the model of a detailed target architecture differ from the current architecture?\n3. What is the role of domain experts in the process of generating suggestions for the target architecture?", "prev_section_summary": "The section discusses the purpose of the noSuccessor and noPredecessor sets in the transformation model of a current architecture, how the model considers applications, services, and business building blocks to generate suggestions for a detailed target architecture, and the process for overruling inappropriate suggestions made by the model for a domain expert. The section also explains the different subsets of building blocks in the transformation model and how the model generates suggestions for a detailed target architecture by following a suggestion and processing all sets of successor relationships.", "section_summary": "The section discusses the use of a Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation in generating suggestions for a target architecture. The model of a detailed target architecture is generated by processing all sets of successor relationships and allowing domain experts to overrule inappropriate suggestions. The suggestions for provided services are generated by checking if they are used by an application that is part of the target architecture or the consuming application has a successor relationship. If there are no successor or predecessor relationships, a manual addition of provided services and their business building blocks in the target architecture is necessary.", "excerpt_keywords": "1. Successor relationships\n2. Target architecture\n3. Dependencies\n4. Suggestions\n5. Provided services\n6. NoSuccessor set\n7. NoPredecessor set\n8. Domain expert\n9. Fairness\n10. Positivity"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "bd8e03bc-e601-4841-824f-ed9aeec49953", "node_type": "4", "metadata": {"page_label": "479", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "3f51f64cdd0021155737e7826c409d607939518f65e164e3ee555f38594170fc"}, "2": {"node_id": "7e4e6d40-6aa8-4450-845d-c01cb3d0879c", "node_type": "1", "metadata": {"page_label": "479", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "087f635b21ff7594f883b7e019bad64e1ca1d37d10959f306b65e18768dc9537"}}, "hash": "217e6b4b19a89b7c20815788fc428cd116acfcf20bca540432079de9f4979f13", "text": "getting more detailed,as all sets of successor relationships are getting processed. A suggestion may\nbe inappropriate for a domain expert she can overrule it by modeling di\ufb00erent\ndetails. The result is a model of a detailed target architecture. At \ufb01rst all ser-vices are transferred to the model of a detailed target architecture. Then the\ndependencies can be added to the model of the detailed target architecture.\nSuggestions for Provided Services.\n1.noSuccessor set: for each provided service in the current architecture check\nif it is used by an application that is part of the target architecture or theconsuming application has a successor relationship.\n(a) If there are any applications it is necessary to check if they still can work\nproperly without consuming the service.\n(b) Otherwise, no information from the current architecture is added to the\ntarget architecture.\n2.noPredecessor set: it is not possible to suggest a detail for the target architec-\nture as there exists no detail in the current architecture. A manual addition of\nprovided services and their business building blocks in the target architecture\nis necessary.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "74e7fd9b-b060-48cf-80ce-3e8c7bd02324": {"__data__": {"id_": "74e7fd9b-b060-48cf-80ce-3e8c7bd02324", "embedding": null, "metadata": {"page_label": "480", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the \"oneToMany\" set in the context of enterprise architecture planning?\n2. How does the \"manyToOne\" set suggest that successors provide services in the target architecture?\n3. What is the role of the domain expert in the enterprise architecture planning process?", "prev_section_summary": "The section discusses the use of a Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation in generating suggestions for a target architecture. The model of a detailed target architecture is generated by processing all sets of successor relationships and allowing domain experts to overrule inappropriate suggestions. The suggestions for provided services are generated by checking if they are used by an application that is part of the target architecture or the consuming application has a successor relationship. If there are no successor or predecessor relationships, a manual addition of provided services and their business building blocks in the target architecture is necessary.", "section_summary": "The section discusses the role of the \"oneToMany\" and \"manyToOne\" sets in enterprise architecture planning, as well as the role of the domain expert in the process. The \"oneToMany\" set suggests that successors provide services in the target architecture, while the \"manyToOne\" set suggests that the successor provides services already provided in the current architecture and provides all services of the other predecessors. The domain expert is responsible for modeling additional services and ensuring that all provided services have been modeled in the target architecture.", "excerpt_keywords": "1. Architecture,\n2. Succession,\n3. Transformation,\n4. Business building blocks,\n5. Services,\n6. Predecessor,\n7. Successor,\n8. OneToMany,\n9. ManyToOne,\n10. ManyToMany."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "4d5cdb46-0352-41e4-9e9c-957f124e0967", "node_type": "4", "metadata": {"page_label": "480", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a5553ea6ed07b8e093c44b85d94740f40d50a4c641b848483c962fff81cd2291"}, "3": {"node_id": "b1239ece-1757-4579-9a39-e084b8c6f1c8", "node_type": "1", "metadata": {"page_label": "480", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "c86ea884c7dafe91d43ae3944367ba7a92fb87647a46f74d5787bf7498718274"}}, "hash": "390a4309d4d50cb9ef7e07d7a8844070cd74628df663f796f1e5936ee9958153", "text": "480 P. Diefenthaler and B. Bauer\n3.oneToMany set:\n(a) If the predecessor is part of onlyCurrentArchitecture all provided services\nof the predecessor, including their business building blocks, are suggested\nto be provided by one of the successor applications.\n(b) Otherwise, all provided services and business building blocks of the pre-\ndecessor are suggested to be provided by one of the successor applications\nor the remaining part of the predecessor in the target architecture.\n4.manyToOne set:\n(a) If the successor is part of onlyTargetArchitecture it is suggested to provide\neach service of its successors, but only one per business building block.\n(b) Otherwise, it is suggested that the successor provides the services already\nprovided in the current architecture, i.e. by itself, and provide all services\nof the other predecessors, but only one per business building block.\n5.manyToMany set: All provided services are suggested to be provided by one of\nthe successors. If more than one predecessor provides a service with the same\nbusiness building block the suggestion is to provide only one service in thetarget architecture with such a business building block. Further suggestions\nwere not identi\ufb01ed as this type represents a complex type of restructuring.\nNevertheless, the domain expert should be supported with information aboutapplications changing business support and assigned customer groups. Fur-\nthermore, information which applications belong to onlyCurrentArchitecture\nand onlyTargetArchitecture needs to be presented to the domain expert.\n6.oneToOne set: all services, including their business building blocks, provided\nby the predecessor are suggested to be provided by the successor.\n7. Furthermore, the domain expert can model additional services or let suggested\nservices be provided by an application that is not a successor of the application\nthat provided it in the current architecture.\n8. For each service information is stored if it is the successor of one or more\nservices in the current architecture. This is necessary to allow a sound trans-\nformation planning [ 9].\nAs a result all provided services have been modeled in the target architecture\nincluding their business building blocks. Furthermore, the information aboutsuccessor", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "b1239ece-1757-4579-9a39-e084b8c6f1c8": {"__data__": {"id_": "b1239ece-1757-4579-9a39-e084b8c6f1c8", "embedding": null, "metadata": {"page_label": "480", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of modeling additional services in the target architecture, and how does it contribute to sound transformation planning?\n2. How does the manyToMany set suggest used services of predecessors to be used by at least one successor, and what is the domain expert's role in this process?\n3. What is the difference between the oneToOne and manyToOne sets, and how do they suggest used services to be used in the target architecture?", "prev_section_summary": "The section discusses the role of the \"oneToMany\" and \"manyToOne\" sets in enterprise architecture planning, as well as the role of the domain expert in the process. The \"oneToMany\" set suggests that successors provide services in the target architecture, while the \"manyToOne\" set suggests that the successor provides services already provided in the current architecture and provides all services of the other predecessors. The domain expert is responsible for modeling additional services and ensuring that all provided services have been modeled in the target architecture.", "section_summary": "The section discusses the purpose of modeling additional services in the target architecture and how it contributes to sound transformation planning. It also explains the difference between the oneToOne and manyToOne sets and their role in suggesting used services to be used in the target architecture. The section also highlights the domain expert's role in suggesting used services and the importance of storing information about successor relationships of the services. Additionally, the section discusses the use of manyToMany, oneToOne, and manyToOne sets to suggest used services and the ability for the domain expert to choose which services are used by multiple successors.", "excerpt_keywords": "1. Succession planning\n2. Service modeling\n3. Business building blocks\n4. Transformation planning\n5. Domain expert\n6. Predecessor\n7. Successor\n8. Many-to-many relationships\n9. One-to-one relationships\n10. Many-to-one relationships"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "4d5cdb46-0352-41e4-9e9c-957f124e0967", "node_type": "4", "metadata": {"page_label": "480", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a5553ea6ed07b8e093c44b85d94740f40d50a4c641b848483c962fff81cd2291"}, "2": {"node_id": "74e7fd9b-b060-48cf-80ce-3e8c7bd02324", "node_type": "1", "metadata": {"page_label": "480", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "390a4309d4d50cb9ef7e07d7a8844070cd74628df663f796f1e5936ee9958153"}}, "hash": "c86ea884c7dafe91d43ae3944367ba7a92fb87647a46f74d5787bf7498718274", "text": "including their business building blocks, provided\nby the predecessor are suggested to be provided by the successor.\n7. Furthermore, the domain expert can model additional services or let suggested\nservices be provided by an application that is not a successor of the application\nthat provided it in the current architecture.\n8. For each service information is stored if it is the successor of one or more\nservices in the current architecture. This is necessary to allow a sound trans-\nformation planning [ 9].\nAs a result all provided services have been modeled in the target architecture\nincluding their business building blocks. Furthermore, the information aboutsuccessor relationships of the services is available.\nSuggestions for Used Services.\n1.manyToMany set: all used services of predecessors are suggested to be used\nby at least one successor. The domain expert can choose if more than one\nsuccessor uses the service of a predecessor.\n2.oneToOne set: all services used by the predecessor are suggested to be used\nby the successor.\n3.manyToOne set: used services of the predecessors are suggested to be also\nused in the target architecture.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "ae7bdbae-9d01-44e6-8f71-00f59971ea39": {"__data__": {"id_": "ae7bdbae-9d01-44e6-8f71-00f59971ea39", "embedding": null, "metadata": {"page_label": "481", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the \"oneToMany set\" in the guided refinement process for enterprise architecture planning?\n2. How does the \"noPredecessor set\" and \"noSuccessor set\" affect the suggested services for an application in the target architecture?\n3. What is the role of the domain expert in the guided refinement process for enterprise architecture planning?", "prev_section_summary": "The section discusses the purpose of modeling additional services in the target architecture and how it contributes to sound transformation planning. It also explains the difference between the oneToOne and manyToOne sets and their role in suggesting used services to be used in the target architecture. The section also highlights the domain expert's role in suggesting used services and the importance of storing information about successor relationships of the services. Additionally, the section discusses the use of manyToMany, oneToOne, and manyToOne sets to suggest used services and the ability for the domain expert to choose which services are used by multiple successors.", "section_summary": "The section discusses the guided refinement process for enterprise architecture planning, specifically focusing on the \"oneToMany set,\" \"noPredecessor set,\" and \"noSuccessor set.\" The purpose of these sets is to suggest services for an application in the target architecture based on the current architecture. The role of the domain expert in this process is to model additional used services for each application. The results of the guided refinement process include a detailed target architecture with provided and used services, and related business building blocks. Gap analysis can be performed again to identify any remaining gaps between the current and target architecture. An action repository is also discussed, which is used to describe possible changes in the transformation path from the current to the target architecture.", "excerpt_keywords": "1. Enterprise Architecture Planning\n2. OneToMany set\n3. Predecessor\n4. Successor\n5. Gaps\n6. Transformation Paths\n7. Action Repository\n8. Abstract actions\n9. Preconditions\n10. Sequencing of actions"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "2e1c4628-7cdd-4f08-8d86-072b77ae9761", "node_type": "4", "metadata": {"page_label": "481", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "4ad5c5d20a0766e3c50e1d4af23dfb8e23882a51c39751faad138e202e6a1fb4"}, "3": {"node_id": "0d9a1aed-8f69-4212-80e7-44708a37f5e7", "node_type": "1", "metadata": {"page_label": "481", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "1d53072bc20a297afec55ff39fe0121626e11660020eef0226cab0537193ef8e"}}, "hash": "0b532f10a66745b59483d59d092b62503179cd65d46c2b21799315ade3ff5410", "text": "From Gaps to Transformation Paths in Enterprise Architecture Planning 481\n4.oneToMany set:\n(a) If the predecessor is part of onlyCurrentArchitecture all used services of\nthe predecessor are suggested to be used by one of the successor applica-\ntions.\n(b) Otherwise, all used services of the predecessor are suggested to be used by\none of the successor applications or the remaining part of the predecessor\nin the target architecture.\n5.noPredecessor set: which services are used by the application need to be mod-\neled manually as no information from the model of the current architecture\nis available.\n6.noSuccessor set: as the application does not exist in the model of the target\narchitecture no information about used services needs to be added to the\ntarget architecture.\n7. Furthermore, the domain expert can model additionally used services for\nevery application.\nResults of the Guided Re\ufb01nement. The result is a model of a detailed target\narchitecture including provided and used services with related business building\nblocks. Consistency checks can be performed on the model to check whetherservices exist which are provided but no longer used by any application. Gap\nanalysis can be performed again and the detailed gaps between the models of\nthe current and target architecture are available.\nWith the results of gap analysis and a detailed current architecture it is\npossible to assist a domain expert in modeling a detailed target architectureby making suggestions how to detail it based on the current architecture. The\nvariety of suggestions that can be provided is limited to the information available\nin the EA model. For example, technical information about the services canbe added to allow more sophisticated suggestions, like to prefer web service\ntechnology for services of applications that have to be build.\n3.5 Creating an Action Repository\nBefore the transformation path from the current to the target architecture can\nbe created, it is necessary to describe possible changes in a way which allows the\nsequencing of actions. This is realized with an action repository where abstractactions are modeled. An abstract action consists of two parts. One part speci\ufb01es\nthe preconditions for an action to be applicable. The other part is", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "0d9a1aed-8f69-4212-80e7-44708a37f5e7": {"__data__": {"id_": "0d9a1aed-8f69-4212-80e7-44708a37f5e7", "embedding": null, "metadata": {"page_label": "481", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of creating an action repository in the Enterprise Architecture (EA) model?\n2. How does the abstract action in the action repository match with the concrete model of the different states?\n3. What is the role of the action repository in the transformation path from the current to the target architecture?", "prev_section_summary": "The section discusses the guided refinement process for enterprise architecture planning, specifically focusing on the \"oneToMany set,\" \"noPredecessor set,\" and \"noSuccessor set.\" The purpose of these sets is to suggest services for an application in the target architecture based on the current architecture. The role of the domain expert in this process is to model additional used services for each application. The results of the guided refinement process include a detailed target architecture with provided and used services, and related business building blocks. Gap analysis can be performed again to identify any remaining gaps between the current and target architecture. An action repository is also discussed, which is used to describe possible changes in the transformation path from the current to the target architecture.", "section_summary": "The section discusses the purpose and role of an action repository in the Enterprise Architecture (EA) model. The action repository is used to model abstract actions that specify preconditions and effects for changes to an architecture model. The actions are described on an abstract level and can be checked for impact if the meta-model of the EA changes. The section also explains how the abstract action matches with the concrete model of the different states through a graph pattern.", "excerpt_keywords": "1. Action repository\n2. Abstract actions\n3. Preconditions\n4. Effect part\n5. Architecture model\n6. Sequencing of actions\n7. Meta-model\n8. Graph pattern\n9. Concrete actions\n10. Concrete entities"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "2e1c4628-7cdd-4f08-8d86-072b77ae9761", "node_type": "4", "metadata": {"page_label": "481", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "4ad5c5d20a0766e3c50e1d4af23dfb8e23882a51c39751faad138e202e6a1fb4"}, "2": {"node_id": "ae7bdbae-9d01-44e6-8f71-00f59971ea39", "node_type": "1", "metadata": {"page_label": "481", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "0b532f10a66745b59483d59d092b62503179cd65d46c2b21799315ade3ff5410"}}, "hash": "1d53072bc20a297afec55ff39fe0121626e11660020eef0226cab0537193ef8e", "text": "the EA model. For example, technical information about the services canbe added to allow more sophisticated suggestions, like to prefer web service\ntechnology for services of applications that have to be build.\n3.5 Creating an Action Repository\nBefore the transformation path from the current to the target architecture can\nbe created, it is necessary to describe possible changes in a way which allows the\nsequencing of actions. This is realized with an action repository where abstractactions are modeled. An abstract action consists of two parts. One part speci\ufb01es\nthe preconditions for an action to be applicable. The other part is the e\ufb00ect part,\nwhich speci\ufb01es the changes to an architecture model if an (abstract) action is\napplied to it.\nThe creation of the action repository is only done once as the actions are\ndescribed on an abstract level. However, if the meta-model of the EA changes\nthe actions in the action repository need to be checked if they are impacted by\nthese changes.\nIn a technical sense the abstract action matches via a graph pattern into the\nconcrete model of the di\ufb00erent states. Concrete actions relate to concrete entities", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "f64a2da0-63e9-404d-9e95-e8d462c4da03": {"__data__": {"id_": "f64a2da0-63e9-404d-9e95-e8d462c4da03", "embedding": null, "metadata": {"page_label": "482", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the difference between atomic and composed actions in the context of enterprise architecture planning?\n2. How does the logical order of abstract actions prevent the creation of loops in the transformation path?\n3. Can you explain how the action repository and transformation model are used to create possible transformation paths in enterprise architecture planning?", "prev_section_summary": "The section discusses the purpose and role of an action repository in the Enterprise Architecture (EA) model. The action repository is used to model abstract actions that specify preconditions and effects for changes to an architecture model. The actions are described on an abstract level and can be checked for impact if the meta-model of the EA changes. The section also explains how the abstract action matches with the concrete model of the different states through a graph pattern.", "section_summary": "The section discusses the concepts of atomic and composed actions in the context of enterprise architecture planning, and how the logical order of abstract actions prevents the creation of loops in the transformation path. It also explains how the action repository and transformation model are used to create possible transformation paths in enterprise architecture planning. The section highlights the importance of modeling abstract actions for shutting down and developing building blocks and abstract actions that take care of the relationships between the building blocks and the attributes of the building blocks. The logical order of abstract actions prevents the creation of loops in the transformation path and ensures that the predecessor services are not shut down until the successor service is developed.", "excerpt_keywords": "1. Architecture model\n2. Abstract actions\n3. Concrete actions\n4. Transformation path\n5. Planning knowledge base\n6. Logical order\n7. Dependencies\n8. Service\n9. Application\n10. Development"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "6de21fe1-d7d1-47c2-be30-c83697669eb5", "node_type": "4", "metadata": {"page_label": "482", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "15672c9e516b1d378e4664126899f48bc6dc97265e02e44fa6195f7c530592fd"}, "3": {"node_id": "5877d42f-f6c7-4b0c-a003-cc67dd788599", "node_type": "1", "metadata": {"page_label": "482", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "5c94160995dc0468555255e891b2069b8f83210ceda577a133c5f8c75e3352b5"}}, "hash": "4b7953c1a33b54ce227322e1ffa98f4bddb8671d8f5af9a123509b1c1c426d63", "text": "482 P. Diefenthaler and B. Bauer\nand relationships in an architecture model and concrete changes to the state of\narchitecture models. The application of a concrete action to an architecturemodel, may enable the application of several other concrete actions.\nAbstract actions are either atomic or composed. An atomic action changes\nexactly one element of either currentArchitecture ortargetArchitecture . Com-\nposed actions are a composition of other actions, regardless if atomic or com-\nposed. To create a transformation path it is necessary to model at least abstract\nactions for shutting down and developing building blocks and abstract actionsthat take care of the relationships between the building blocks and the attributes\nof the building blocks.\nLogical Order of Abstract Actions. The abstract actions are modeled in\na logical order, which means that it is only possible to apply the action if the\npreceding actions were already applied. For example, it is not possible to changethe dependencies from a service to its successor service if it has not yet been\nbuilt. Furthermore, it may be necessary to build the application \ufb01rst to allow the\ncreation of a new service. After the dependencies of a service have been changed\nto a successor it is possible to shutdown the service.\nIf all services of an application have been shutdown it is possible to shutdown\nthe application. The logical order prevents the creation of loops in the transfor-\nmation path, i.e. to shutdown and create the same application several times. It\nmay be the case that it is not necessary to enact the develop application action.\nFor example, if a service which has to be developed for an application that\nalready exists. In this case it is not necessary to develop that application again\nsince it already exists in the current architecture. The logical order prevents theshutdown of the predecessor services, until the successor service is developed.\n3.6 Creating the Transformation Path\nWith the action repository, the transformation model, the models of current\narchitecture and target architecture at hand it is possible to start the creationof possible transformation paths.\nWe derive all applicable concrete actions by checking which preconditions of\nabstract actions match in\nplanningKnowledgeBase :={transf", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "5877d42f-f6c7-4b0c-a003-cc67dd788599": {"__data__": {"id_": "5877d42f-f6c7-4b0c-a003-cc67dd788599", "embedding": null, "metadata": {"page_label": "482", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using a breadth search versus a depth search in creating transformation paths in enterprise architecture planning?\n2. How does the action repository and the transformation model aid in the creation of possible transformation paths in enterprise architecture planning?\n3. What is the logical order for developing a new service in enterprise architecture planning, and how does it prevent the shutdown of predecessor services?", "prev_section_summary": "The section discusses the concepts of atomic and composed actions in the context of enterprise architecture planning, and how the logical order of abstract actions prevents the creation of loops in the transformation path. It also explains how the action repository and transformation model are used to create possible transformation paths in enterprise architecture planning. The section highlights the importance of modeling abstract actions for shutting down and developing building blocks and abstract actions that take care of the relationships between the building blocks and the attributes of the building blocks. The logical order of abstract actions prevents the creation of loops in the transformation path and ensures that the predecessor services are not shut down until the successor service is developed.", "section_summary": "The section discusses the use of a breadth search versus a depth search in creating transformation paths in enterprise architecture planning. It also explains how the action repository and the transformation model aid in the creation of possible transformation paths. The logical order for developing a new service in enterprise architecture planning is also discussed, as well as the process of creating the transformation path using the action repository, the transformation model, and the models of the current and target architectures.", "excerpt_keywords": "1. Transformation model\n2. Action repository\n3. Planning knowledge base\n4. Current architecture\n5. Target architecture\n6. Concrete actions\n7. Abstract actions\n8. Preconditions\n9. Breadth search\n10. Depth search"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "6de21fe1-d7d1-47c2-be30-c83697669eb5", "node_type": "4", "metadata": {"page_label": "482", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "15672c9e516b1d378e4664126899f48bc6dc97265e02e44fa6195f7c530592fd"}, "2": {"node_id": "f64a2da0-63e9-404d-9e95-e8d462c4da03", "node_type": "1", "metadata": {"page_label": "482", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "4b7953c1a33b54ce227322e1ffa98f4bddb8671d8f5af9a123509b1c1c426d63"}}, "hash": "5c94160995dc0468555255e891b2069b8f83210ceda577a133c5f8c75e3352b5", "text": "which has to be developed for an application that\nalready exists. In this case it is not necessary to develop that application again\nsince it already exists in the current architecture. The logical order prevents theshutdown of the predecessor services, until the successor service is developed.\n3.6 Creating the Transformation Path\nWith the action repository, the transformation model, the models of current\narchitecture and target architecture at hand it is possible to start the creationof possible transformation paths.\nWe derive all applicable concrete actions by checking which preconditions of\nabstract actions match in\nplanningKnowledgeBase :={transf ormationM odel \u222acurrentArchitecture \u222a\ntargetArchitecture }\nThis corresponds to a breadth search of applicable actions for a possible change\nfrom the current towards the target architecture. If a concrete action is applied to\nplanningKnowledgeBase it changes the state of the planningKnowledgeBase .I n\ncontrast if we apply a depth search we receive a transformation path changing the\nEA in a sequence of concrete actions from the current to the target architecture.\nIf no such transformation path exists the more exhaustive breadth search canbe omitted and we are informed that no transformation path was found. By\napplying the breadth search on each state recursively and we get the whole state\nspace.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "5eb87bf6-d65a-494f-9093-4eb198babbc3": {"__data__": {"id_": "5eb87bf6-d65a-494f-9093-4eb198babbc3", "embedding": null, "metadata": {"page_label": "483", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation?\n2. How does the selection process for choosing concrete actions in the transformation path enhancement work?\n3. What is the role of master data management in the development of applications in an organization's IT landscape?", "prev_section_summary": "The section discusses the use of a breadth search versus a depth search in creating transformation paths in enterprise architecture planning. It also explains how the action repository and the transformation model aid in the creation of possible transformation paths. The logical order for developing a new service in enterprise architecture planning is also discussed, as well as the process of creating the transformation path using the action repository, the transformation model, and the models of the current and target architectures.", "section_summary": "The section discusses the use of a Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation in enterprise architecture planning. The network is used to determine all possible transformation paths from the current to the target architecture and to select concrete actions for the transformation path enhancement process. The section also discusses the role of master data management in the development of applications in an organization's IT landscape and provides a use case for the introduction of master data management in the research and development division of an organization. The use case shows a part of the model of the current architecture of the organization's IT landscape and the development of a master data management system to provide services to other applications.", "excerpt_keywords": "Enterprise architecture planning, state space, transformation paths, concrete actions, development costs, maintenance costs, benefits, risks, resource constraints, master data management, development master data management, research and development division, IT landscape, applications, data consistency, redundancy, outdated data, organization, model, current architecture, development master data management system, services, DevManager, product planning tool, quality tests planning tool, virtual quality test result database, master data."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "0de2524a-66fb-45c2-8a35-7ad0f2e66e36", "node_type": "4", "metadata": {"page_label": "483", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "44b70d9dd3875a1574c410efe32c278ab0e1181c77c9ec481e2b18c5e3b978a0"}, "3": {"node_id": "6672141f-5ff2-44e3-95d9-49dd3bfd4b97", "node_type": "1", "metadata": {"page_label": "483", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "1770fe428768b7656b5f49068ca32ad8c5a31fa19be597cf100fc5ad43be5b73"}}, "hash": "81e51deeb46c7f596fbd7220773cfa5206fcfb0cb511d350bfa751c9a177e04d", "text": "From Gaps to Transformation Paths in Enterprise Architecture Planning 483\nWith the state space it is possible to determine all possible transformation\npaths from the current to the target architecture. By selecting concrete actionswe create the transformation path, change the planningKnowledgeBase and get\neach time a list of concrete actions which we now can apply. When the transfor-\nmation path is complete, i.e. all necessary changes have been applied, no furtheractions are applicable and the transformation path is saved. If gaps are not to\nbe closed it is possible to stop the creation of the transformation path.\nThe selection process for choosing concrete actions can be enhanced by pro-\nviding development costs for proposed applications and services, and mainte-\nnance costs for applications and services which are to be retired. Furthermore,\nthe consideration of desired bene\ufb01ts, anticipated risks and resource constraintscould be considered if available to allow for a weighting of favorable sequences\nof actions.\n4 Use Case - Development Master Data Management\nIn the past, applications were often developed to address the speci\ufb01c business\nneeds that a part of the organization had at a certain moment. However, consid-ering the whole enterprise it is not e\ufb00ective to store redundant data in several\napplications as this increases the risk of outdated and inconsistent data. This\nis the basis for the master data management (MDM) challenge [ 26]. In our use\ncase we show a typical (and simpli\ufb01ed) example for the introduction of master\ndata management in the research and development division of an organization.Figure 1shows a part of the model of the current architecture of the organi-\nzation\u2019s IT landscape. There has already been placed a development master\ndata management (DMDM) system in the organization which provides services(MasterData\nv1 and v2) to other applications. However, not all existing appli-\ncations use the master data provided by DMDM: the application DevManager\nprovides similar data that is still used by existing applications such as the prod-uct planning tool and the quality tests planning tool. Other applications such\nas the virtual quality test result database store the master", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "6672141f-5ff2-44e3-95d9-49dd3bfd4b97": {"__data__": {"id_": "6672141f-5ff2-44e3-95d9-49dd3bfd4b97", "embedding": null, "metadata": {"page_label": "483", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What are the current gaps in the organization's IT landscape in terms of data management and planning tools?\n2. How does the proposed target architecture aim to address these gaps and unify the functionality of different applications?\n3. What are the specific tools and systems that will be used in the target architecture to manage master data and quality tests?", "prev_section_summary": "The section discusses the use of a Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation in enterprise architecture planning. The network is used to determine all possible transformation paths from the current to the target architecture and to select concrete actions for the transformation path enhancement process. The section also discusses the role of master data management in the development of applications in an organization's IT landscape and provides a use case for the introduction of master data management in the research and development division of an organization. The use case shows a part of the model of the current architecture of the organization's IT landscape and the development of a master data management system to provide services to other applications.", "section_summary": "The section discusses the current gaps in an organization's IT landscape in terms of data management and planning tools. The proposed target architecture aims to address these gaps by unifying the functionality of different applications and using a development master data management (DMDM) system to manage master data and quality tests. The specific tools and systems that will be used in the target architecture include the DMDM system, a planning tool that includes planning for the product and quality tests, and a quality test assistance and result management tool. The section also mentions the existence of different applications for modifying products and managing quality tests.", "excerpt_keywords": "1. Master data management\n2. Development master data management\n3. Application development\n4. Product planning\n5. Quality tests planning\n6. Quality test result management\n7. Virtual quality test result database\n8. Product modification\n9. Product classes\n10. Unified functionality"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "0de2524a-66fb-45c2-8a35-7ad0f2e66e36", "node_type": "4", "metadata": {"page_label": "483", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "44b70d9dd3875a1574c410efe32c278ab0e1181c77c9ec481e2b18c5e3b978a0"}, "2": {"node_id": "5eb87bf6-d65a-494f-9093-4eb198babbc3", "node_type": "1", "metadata": {"page_label": "483", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "81e51deeb46c7f596fbd7220773cfa5206fcfb0cb511d350bfa751c9a177e04d"}}, "hash": "1770fe428768b7656b5f49068ca32ad8c5a31fa19be597cf100fc5ad43be5b73", "text": "1shows a part of the model of the current architecture of the organi-\nzation\u2019s IT landscape. There has already been placed a development master\ndata management (DMDM) system in the organization which provides services(MasterData\nv1 and v2) to other applications. However, not all existing appli-\ncations use the master data provided by DMDM: the application DevManager\nprovides similar data that is still used by existing applications such as the prod-uct planning tool and the quality tests planning tool. Other applications such\nas the virtual quality test result database store the master data themselves and\nare not connected to DMDM. For the modi\ufb01cation of products (from one testto another) there exist two applications for the di\ufb00erent product classes the\norganization provides to their customers. Additionally, applications to plan the\nproduct, the quality tests and store the results that have been gathered duringthe (physical or virtual) quality tests, exist. In the model of the target architec-\nture the functionality in the di\ufb00erent applications shall be united and all other\ntools will use the data provided by DMDM. There will be only one planning tool\nthat includes planning for the product as well as the quality tests. All quality\ntests (including the results) will be managed by one quality test assistance andresult management tool (cf. Figure 2).\nPlease note that Figs. 1and2already contain the services, which may not\nbe considered in the \ufb01rst place for planning purposes.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "517ceb00-a21c-43fd-a3e1-999e2a837bc8": {"__data__": {"id_": "517ceb00-a21c-43fd-a3e1-999e2a837bc8", "embedding": null, "metadata": {"page_label": "484", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the current architecture of master data management in the DevManager system, and what are the key components involved in it?\n2. What are the target architecture components of master data management in the DevManager system, and how do they differ from the current architecture?\n3. What are the successor relationships between the various components of master data management in the DevManager system, and how do they relate to each other in the transformation model?", "prev_section_summary": "The section discusses the current gaps in an organization's IT landscape in terms of data management and planning tools. The proposed target architecture aims to address these gaps by unifying the functionality of different applications and using a development master data management (DMDM) system to manage master data and quality tests. The specific tools and systems that will be used in the target architecture include the DMDM system, a planning tool that includes planning for the product and quality tests, and a quality test assistance and result management tool. The section also mentions the existence of different applications for modifying products and managing quality tests.", "section_summary": "The section discusses the current and target architectures of master data management in the DevManager system, as well as the successor relationships between the various components of master data management in the transformation model. The section also mentions the key components involved in the current architecture, which include DevManager, Product Planning Tool, Quality tests planning tool, Physical quality test assistance tool, Physical quality test result database, Virtual quality test result database, Product class A assistance database, Product class B assistance database, QueryDev v1, MasterData v1 and v2. The target architecture includes Development master data management system (DMDM), Product and Quality test planning tool, Quality test assistance and result management tool, Product modification assistance database, MasterData v3 and PlanningData v1. The section also mentions the successor relationships between the various components of master data management in the transformation model, including the successor relationships between the services.", "excerpt_keywords": "1. Master data management, 2. Architecture, 3. Gap analysis, 4. Development master data management system (DMDM), 5. Product planning tool, 6. Quality tests planning tool, 7. Physical quality test assistance tool, 8. Virtual quality test result database, 9. Product modification assistance database, 10. Successor relationships."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "f9d2322a-f237-46b4-939a-b90f81d0e922", "node_type": "4", "metadata": {"page_label": "484", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "ecc43c6e256acb254beef7108c49b128ae23445c3a0e57659376c374f0094682"}, "3": {"node_id": "32613ead-e644-4000-97ff-2bcad3244831", "node_type": "1", "metadata": {"page_label": "484", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "7d253b9f0eae452661d78fe063159247e0fb13b97fe96af5507af94763eeb6e5"}}, "hash": "7c1f1267fcb112fff88743441d95a64d1a90c606163be178e9a523436c9a7c54", "text": "484 P. Diefenthaler and B. Bauer\nDevManager\nQuality tests\nplanning  tool Product planning \ntool\nDevelopment master\ndata management\nsystem (DMDM) Physical quality \ntest assistance \ntool Physical quality \ntest result\ndatabaseVirtual quality \ntest result\ndatabase\nProduct class A \nassistance\ndatabase\nProduct class B \nassistance\ndatabase\nFig. 1. Master data management: current architecture.\nDevelopment master\ndata management\nsystem (DMDM) Product and Quality \ntest planning toolQuality test assistance\nand result management\ntool\nProduct modification\nassistance database\nFig. 2. Master data management: target architecture.\nSolution Applied to the Use Case\nAt \ufb01rst currentArchitecture and targetArchitecture are created by modelling both\narchitectures. Applying gap analysis it is possible to derive that onlyCurrentAr-\nchitecture contains: DevManager, Product Planning Tool, Quality tests planning\ntool, Physical quality test assistance tool, Physical quality test result database,\nVirtual quality test result database, Product class A assistance database, Prod-\nuct class B assistance database, QueryDev v1, MasterData v1 and v2.\nThe set stable contains only Development master data management system\n(DMDM) whereas onlyTargetArchitecture contains Product and Quality test\nplanning tool, Quality test assistance and result management tool, Product mod-\ni\ufb01cation assistance database, MasterData v3 and PlanningData v1.\nWithin the transformation model information on the successor relationships\nis kept: Product planning tool and Quality tests planning tool have the same suc-\ncessor (Product and Quality test planning tool). Physical quality test assistance\ntool, Physical quality test result database and Virtual quality test result data-base have the Quality test assistance and result management tool as a common\nsuccessor. DMDM is a successor of itself, which is in accordance with [ 7], and\nDevManager has no successor. Product modi\ufb01cation assistance database is thesuccessor of Product class A assistance database and product class B assistance\ndatabase.\nRegarding the services the following successor relationships are contained in\nthe transformation model:", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "32613ead-e644-4000-97ff-2bcad3244831": {"__data__": {"id_": "32613ead-e644-4000-97ff-2bcad3244831", "embedding": null, "metadata": {"page_label": "484", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the successor relationship between the Product planning tool and Quality tests planning tool?\n2. What is the successor relationship between the Physical quality test assistance tool, Physical quality test result database, and Virtual quality test result database?\n3. What is the successor relationship between the Product modification assistance database and the Product class A assistance database and product class B assistance database?", "prev_section_summary": "The section discusses the current and target architectures of master data management in the DevManager system, as well as the successor relationships between the various components of master data management in the transformation model. The section also mentions the key components involved in the current architecture, which include DevManager, Product Planning Tool, Quality tests planning tool, Physical quality test assistance tool, Physical quality test result database, Virtual quality test result database, Product class A assistance database, Product class B assistance database, QueryDev v1, MasterData v1 and v2. The target architecture includes Development master data management system (DMDM), Product and Quality test planning tool, Quality test assistance and result management tool, Product modification assistance database, MasterData v3 and PlanningData v1. The section also mentions the successor relationships between the various components of master data management in the transformation model, including the successor relationships between the services.", "section_summary": "The section discusses the successor relationships between various tools and databases in the context of enterprise architecture planning. The successor relationships are used to determine the order in which these tools and databases are used in the planning process. The section also mentions the transformation model, which contains information about the successor and predecessor relationships between various services. The action repository is used to develop these services in the correct order. The key topics and entities discussed in the section include Product planning tool, Quality tests planning tool, Physical quality test assistance tool, Physical quality test result database, Virtual quality test result database, Product modification assistance database, Product class A assistance database, Product class B assistance database, MasterData v1, MasterData v2, MasterData v3, QueryDev v1, PlanningData v1, and the transformation model.", "excerpt_keywords": "1. Successor relationships\n2. Transformation model\n3. Action repository\n4. MasterData\n5. Product planning tool\n6. Quality tests planning tool\n7. Physical quality test assistance tool\n8. Physical quality test result database\n9. Virtual quality test result data-base\n10. Product modification assistance database"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "f9d2322a-f237-46b4-939a-b90f81d0e922", "node_type": "4", "metadata": {"page_label": "484", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "ecc43c6e256acb254beef7108c49b128ae23445c3a0e57659376c374f0094682"}, "2": {"node_id": "517ceb00-a21c-43fd-a3e1-999e2a837bc8", "node_type": "1", "metadata": {"page_label": "484", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "7c1f1267fcb112fff88743441d95a64d1a90c606163be178e9a523436c9a7c54"}}, "hash": "7d253b9f0eae452661d78fe063159247e0fb13b97fe96af5507af94763eeb6e5", "text": "Product planning tool and Quality tests planning tool have the same suc-\ncessor (Product and Quality test planning tool). Physical quality test assistance\ntool, Physical quality test result database and Virtual quality test result data-base have the Quality test assistance and result management tool as a common\nsuccessor. DMDM is a successor of itself, which is in accordance with [ 7], and\nDevManager has no successor. Product modi\ufb01cation assistance database is thesuccessor of Product class A assistance database and product class B assistance\ndatabase.\nRegarding the services the following successor relationships are contained in\nthe transformation model: MasterData\nv3 is a successor of MasterData v1 and\nv2. The QueryDev v1 has no successor and PlanningData v1 has no predecessor.\nBased upon this information the action repository can show that it is possibleto develop MasterData\nv3 in the \ufb01rst place or one of the successor applications.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "41f6e3be-c7e8-483d-a345-4b16404411de": {"__data__": {"id_": "41f6e3be-c7e8-483d-a345-4b16404411de", "embedding": null, "metadata": {"page_label": "485", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the document \"From Gaps to Transformation Paths in Enterprise Architecture Planning\"?\n2. What is the process for deriving gaps between the models of a current and target architecture for planning purposes?\n3. How does the action repository aid in the creation of possible transformation paths?", "prev_section_summary": "The section discusses the successor relationships between various tools and databases in the context of enterprise architecture planning. The successor relationships are used to determine the order in which these tools and databases are used in the planning process. The section also mentions the transformation model, which contains information about the successor and predecessor relationships between various services. The action repository is used to develop these services in the correct order. The key topics and entities discussed in the section include Product planning tool, Quality tests planning tool, Physical quality test assistance tool, Physical quality test result database, Virtual quality test result database, Product modification assistance database, Product class A assistance database, Product class B assistance database, MasterData v1, MasterData v2, MasterData v3, QueryDev v1, PlanningData v1, and the transformation model.", "section_summary": "The section discusses a document titled \"From Gaps to Transformation Paths in Enterprise Architecture Planning\" which outlines a process for deriving gaps between the models of a current and target architecture for planning purposes. The process involves using a set theoretic description to reuse existing information and aid in the detailing of the model of the target architecture. An action repository is also used to create possible transformation paths, which are sequences of actions. The solution considers a domain expert as an important part of the activities and assists her in the decision making process. The document also discusses limitations and requirements regarding the meta-model.", "excerpt_keywords": "Enterprise Architecture Planning, MasterData v3, dependencies, successor relationships, gap analysis, transformation model, action repository, business building blocks, meta-model, EAM approach, applications, services."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "0b6c8e68-14a4-482b-a5a8-0e70a36ba52d", "node_type": "4", "metadata": {"page_label": "485", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "0243dbf30e08de942f6fc080f3e64eedaecc3d1d0c734f189cb27f92d1ef18c6"}, "3": {"node_id": "f6bf074d-9de7-49e9-81be-78ff8f27d9b1", "node_type": "1", "metadata": {"page_label": "485", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "1bdc619de56fe859600b6797b824846005b3614712b6f57aa4fc23e1ed20aac3"}}, "hash": "723df848cffc6f2a1c2e88b1a8a302a0651ef07eb8f2aa776b45f1b7b82fd633", "text": "From Gaps to Transformation Paths in Enterprise Architecture Planning 485\nIf for example as the \ufb01rst action the development of MasterData v3 is selected\nit is possible to take care of the dependencies of applications to the predecessorsof the service. After removing the dependencies and creating the new ones to the\nsuccessor service it is possible to shutdown the predecessors. The development\nof the new applications are to be selected as the next steps in the transformationpath. The remaining actions are not described in detail, however their sequence\nis constrained by the logical order of the abstract actions.\n5 Discussion\nThe discussion is divided into two parts. At \ufb01rst we discuss the results of solution\nand its application to the use case. After that, the limitations of the solution arepresented.\nThe solution describes how it is possible to derive gaps between the models\nof a current and target architecture for planning purposes using a set theoreticdescription. With the gaps at hand and information regarding the successor\nrelationships of elements the solution reuses existing information to aid in the\ndetailing the model of the target architecture. Afterwards, an action repositoryaids in the creation of possible transformation paths, which are sequences of\nactions. Overall, the solution considers a domain expert as an important part of\nthe activities and assists her in the decision making process.\nCreating suggestions for detailing the model of a target architecture is only\npossible if business building blocks are available. However, the mechanism ofgap analysis, the transformation model and the creation of transformation paths\nusing the action repository are not impacted by this limitation.\nFurthermore, requirements regarding the meta-model are posed by the solu-\ntion. If the EAM approach does not concern application architectures, and as a\nconsequence the models of applications and their dependencies to services, the\nsolution would in its current shape not be suitable. However, the mechanisms asdescribed in the solution can be adapted to aid in the modelling and creation of\ntransformation paths which address the concerns of the stakeholders. From our\npoint of view, applications and their provided services are an important part ofan EA.\nCurrently, we create the connection of the models of the current and tar-\nget architecture manually, which is prone", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "f6bf074d-9de7-49e9-81be-78ff8f27d9b1": {"__data__": {"id_": "f6bf074d-9de7-49e9-81be-78ff8f27d9b1", "embedding": null, "metadata": {"page_label": "485", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What are the limitations of the current EAM approach in addressing the concerns of stakeholders related to applications and their dependencies on services?\n2. How can the mechanisms described in the solution be adapted to aid in the modelling and creation of transformation paths that take technology architecture aspects into account?\n3. What are the advantages of using semantic web technologies for formalizing information sources in the context of EA planning?", "prev_section_summary": "The section discusses a document titled \"From Gaps to Transformation Paths in Enterprise Architecture Planning\" which outlines a process for deriving gaps between the models of a current and target architecture for planning purposes. The process involves using a set theoretic description to reuse existing information and aid in the detailing of the model of the target architecture. An action repository is also used to create possible transformation paths, which are sequences of actions. The solution considers a domain expert as an important part of the activities and assists her in the decision making process. The document also discusses limitations and requirements regarding the meta-model.", "section_summary": "The section discusses the limitations of the current EAM approach in addressing the concerns of stakeholders related to applications and their dependencies on services. It proposes the use of semantic web technologies for formalizing information sources in the context of EA planning, which has advantages such as a formal, unambiguous model and the possibility of reasoning and consistency checking. The section also mentions the importance of applications and their provided services in EA and the need for transformation paths that take technology architecture aspects into account.", "excerpt_keywords": "1. Enterprise Architecture Modeling\n2. Application Architecture\n3. Service Dependencies\n4. Transformation Paths\n5. Semantic Web Technologies\n6. Formal Modeling\n7. Reasoning\n8. Consistency Checking\n9. Knowledge Base\n10. EAM Approach"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "0b6c8e68-14a4-482b-a5a8-0e70a36ba52d", "node_type": "4", "metadata": {"page_label": "485", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "0243dbf30e08de942f6fc080f3e64eedaecc3d1d0c734f189cb27f92d1ef18c6"}, "2": {"node_id": "41f6e3be-c7e8-483d-a345-4b16404411de", "node_type": "1", "metadata": {"page_label": "485", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "723df848cffc6f2a1c2e88b1a8a302a0651ef07eb8f2aa776b45f1b7b82fd633"}}, "hash": "1bdc619de56fe859600b6797b824846005b3614712b6f57aa4fc23e1ed20aac3", "text": "are posed by the solu-\ntion. If the EAM approach does not concern application architectures, and as a\nconsequence the models of applications and their dependencies to services, the\nsolution would in its current shape not be suitable. However, the mechanisms asdescribed in the solution can be adapted to aid in the modelling and creation of\ntransformation paths which address the concerns of the stakeholders. From our\npoint of view, applications and their provided services are an important part ofan EA.\nCurrently, we create the connection of the models of the current and tar-\nget architecture manually, which is prone to errors and time consuming. Themodel of the target architecture does currently not consider information which\ntransformation paths, taking technology architecture aspects into account, are\npossible.\n6 Proposed Technical Realization\nUsing semantic web technologies for formalizing information sources yields a\nnumber of advantages, starting with having a formal, unambiguous model to\nthe possibilities of reasoning and consistency checking. The knowledge base", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "e12b8d47-e893-40c2-af6f-cf52cf874035": {"__data__": {"id_": "e12b8d47-e893-40c2-af6f-cf52cf874035", "embedding": null, "metadata": {"page_label": "486", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using a graph transformation approach for creating transformation paths in enterprise architecture planning?\n2. What is the difference between using SPARQL and a more sophisticated graph transformation approach for detailing the model of the target architecture?\n3. What is the proposed technical realization for modeling (M2M) transformation of ontologies to a model interpretable for a graph transformation approach?", "prev_section_summary": "The section discusses the limitations of the current EAM approach in addressing the concerns of stakeholders related to applications and their dependencies on services. It proposes the use of semantic web technologies for formalizing information sources in the context of EA planning, which has advantages such as a formal, unambiguous model and the possibility of reasoning and consistency checking. The section also mentions the importance of applications and their provided services in EA and the need for transformation paths that take technology architecture aspects into account.", "section_summary": "The section discusses the use of a graph transformation approach for creating transformation paths in enterprise architecture planning. It compares the usage of SPARQL and a more sophisticated graph transformation approach for detailing the model of the target architecture. The proposed technical realization for modeling (M2M) transformation of ontologies to a model interpretable for a graph transformation approach is discussed, along with related work in the field of enterprise architecture management. The section mentions the use of GROOVE as a mature graph transformation tool and the proposed solution requires a model to model (M2M) transformation of the ontologies to a model which is interpretable for a graph transformation approach.", "excerpt_keywords": "Enterprise Architecture Management, gap analysis, ontologies, graph transformation, Rule Interchange Format, GROOVE, modeling and modeling transformation, SPARQL, EA models, target architecture, transformation paths, ontology transformation, interpretable graph transformation, technical realization, related work, University of Oldenburg, Quasar Enterprise approach, technical standard, The Open Group."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "87f7dab9-8506-4d0c-9b70-4cbc0003cd75", "node_type": "4", "metadata": {"page_label": "486", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "62c283b3781b3eb1252f949a458c78ef86493d1e5524328b881d9b326d7a80ac"}, "3": {"node_id": "63c50146-9ae5-4fbf-bd1f-92f1f0bc0255", "node_type": "1", "metadata": {"page_label": "486", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "aa808aa11f3f8ea3167abcbdd5d5b4ef7c4c1ccff0d42217dd0c874d34d50cea"}}, "hash": "94dd88755af7df284a9baf69633a90eeb66fa1c98cf73ca61bf47ee7bf86ccaf", "text": "486 P. Diefenthaler and B. Bauer\ncontaining, the current and target EA models, as well as the transformation\nmodel, can be consulted at run time by humans as well as by applications.\nIdentifying gaps can be realized using standard tools like Prot\u00b4 eg\u00b4e3for mod-\neling and OWLDi\ufb004for comparing the modeled EAs. For detailing the model of\nthe target architecture we suggest the usage of SPARQL as it allows queryingand adding information in a semi-automated manner.\nRegarding the creation of transformation paths we suggest to use a more\nsophisticated graph transformation approach, as it provides the expressivenessnecessary for the creation of transformation paths. This requirement exceeds\nthe current capabilities of SPARQL. A promising World Wide Web Consor-\ntium standard is the Rule Interchange Format\n5(RIF), which initial purpose\nwas the exchange of rules. The second edition of RIF provides an action lan-\nguage which can be used to express the actions necessary for transformation\npath planning. However, we were not able to test the proposed solution as no\nfree implementations are available yet. Therefore, we propose to use a mature\ngraph transformation tool like GROOVE6.\nHowever, this proposed technical realization requires a model to model (M2M)\ntransformation of the ontologies to a model which is interpretable for a graph\ntransformation approach.\n7 Related Work\nIn this section related work is introduced. As a starting point the technical report\u2018On the state of the Art in Enterprise Architecture Management Literature\u2019 [ 8]\nwas taken, as they consider gap (delta) analysis as part of the di\ufb00erent EAM\napproaches. Besides the listed approaches in the technical report an approach\nfrom the University of Oldenburg and a technical standard from The Open Groupwas identi\ufb01ed as relate work.\n7.1 University of Oldenburg\nThe Institute for Information Technology of the University of Oldenburg presents\na tool supported approach for performing a gap analysis on a current and ideal\nlandscape [ 27]. The approach is tightly coupled to the Quasar Enterprise app-\nroach, which can be used to", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "63c50146-9ae5-4fbf-bd1f-92f1f0bc0255": {"__data__": {"id_": "63c50146-9ae5-4fbf-bd1f-92f1f0bc0255", "embedding": null, "metadata": {"page_label": "486", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the gap analysis approach presented by the Institute for Information Technology at the University of Oldenburg in relation to the Quasar Enterprise app-roach?\n2. How does the suggested procedure for performing a gap analysis by the University of Oldenburg differ from other approaches identified in the technical report?\n3. What is the role of the Open Group technical standard in relation to the gap analysis approaches discussed in the technical report?", "prev_section_summary": "The section discusses the use of a graph transformation approach for creating transformation paths in enterprise architecture planning. It compares the usage of SPARQL and a more sophisticated graph transformation approach for detailing the model of the target architecture. The proposed technical realization for modeling (M2M) transformation of ontologies to a model interpretable for a graph transformation approach is discussed, along with related work in the field of enterprise architecture management. The section mentions the use of GROOVE as a mature graph transformation tool and the proposed solution requires a model to model (M2M) transformation of the ontologies to a model which is interpretable for a graph transformation approach.", "section_summary": "The section discusses the gap analysis approach presented by the Institute for Information Technology at the University of Oldenburg in relation to the Quasar Enterprise app-roach. The approach is tightly coupled to the Quasar Enterprise app-roach and involves modeling the current and ideal application landscapes to generate a list of actions that would result in the ideal landscape. The section also identifies other approaches to gap analysis, including an approach from the University of Oldenburg and a technical standard from The Open Group. The Open Group technical standard is mentioned as being related to the gap analysis approaches discussed in the technical report.", "excerpt_keywords": "gap analysis, service-oriented application landscapes, University of Oldenburg, Quasar Enterprise approach, modeling, current application landscape, ideal landscape, actions, service components, service operations, business objects."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "87f7dab9-8506-4d0c-9b70-4cbc0003cd75", "node_type": "4", "metadata": {"page_label": "486", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "62c283b3781b3eb1252f949a458c78ef86493d1e5524328b881d9b326d7a80ac"}, "2": {"node_id": "e12b8d47-e893-40c2-af6f-cf52cf874035", "node_type": "1", "metadata": {"page_label": "486", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "94dd88755af7df284a9baf69633a90eeb66fa1c98cf73ca61bf47ee7bf86ccaf"}}, "hash": "aa808aa11f3f8ea3167abcbdd5d5b4ef7c4c1ccff0d42217dd0c874d34d50cea", "text": "8]\nwas taken, as they consider gap (delta) analysis as part of the di\ufb00erent EAM\napproaches. Besides the listed approaches in the technical report an approach\nfrom the University of Oldenburg and a technical standard from The Open Groupwas identi\ufb01ed as relate work.\n7.1 University of Oldenburg\nThe Institute for Information Technology of the University of Oldenburg presents\na tool supported approach for performing a gap analysis on a current and ideal\nlandscape [ 27]. The approach is tightly coupled to the Quasar Enterprise app-\nroach, which can be used to develop service-oriented application landscapes.\nIn order to be able to perform their gap analysis it is necessary to model the\ncurrent application landscape consisting of current components, current services,current operations and business objects. The ideal landscape is modeled with\nideal components, ideal services, ideal operations and domains. Based on these\ntwo models the tool is capable to generate a list of actions that would, if all wereapplied, result in the ideal landscape. Within the paper the suggested procedure\n3http://protege.stanford.edu/\n4http://krizik.felk.cvut.cz/km/owldi\ufb00/\n5http://www.w3.org/TR/2013/NOTE-rif-overview-20130205/\n6http://groove.cs.utwente.nl/", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "117b0747-5d77-4e73-a31a-003d96ed1e86": {"__data__": {"id_": "117b0747-5d77-4e73-a31a-003d96ed1e86", "embedding": null, "metadata": {"page_label": "487", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What are the limitations of using gap analysis for selecting actions in Enterprise Architecture Planning?\n2. How does the Strategic IT Management approach use gap analysis to detect differences between the current and target architecture?\n3. What are the possible actions that can be considered to close gaps identified through gap analysis in Enterprise Architecture Planning?", "prev_section_summary": "The section discusses the gap analysis approach presented by the Institute for Information Technology at the University of Oldenburg in relation to the Quasar Enterprise app-roach. The approach is tightly coupled to the Quasar Enterprise app-roach and involves modeling the current and ideal application landscapes to generate a list of actions that would result in the ideal landscape. The section also identifies other approaches to gap analysis, including an approach from the University of Oldenburg and a technical standard from The Open Group. The Open Group technical standard is mentioned as being related to the gap analysis approaches discussed in the technical report.", "section_summary": "The section discusses the limitations of using gap analysis for selecting actions in Enterprise Architecture Planning (EAP). Gap analysis requires a detailed level of description when modeling both landscapes, and the data necessary to perform gap analysis on the entire application landscape on a detailed level considering operations is overwhelming. The Strategic IT Management approach uses gap analysis to detect differences between the current and target architecture, and possible actions to close the gaps identified through gap analysis are considered. The actions range from introducing a new application, adding or reducing functionality of an existing application, changing or adding services to the shut down of applications and services. The section also mentions the ArchiMate Implementation and Migration Extension, which introduces new concepts and modeling capabilities for EAP.", "excerpt_keywords": "1. ArchiMate\n2. Implementation\n3. Migration\n4. Extension\n5. Architecture\n6. Planning\n7. Gap analysis\n8. Process support maps\n9. Services\n10. Information objects"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "672b007c-4b9d-4463-a5d6-c4e1be707ef7", "node_type": "4", "metadata": {"page_label": "487", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "06a251c9de1a6d9955e088165ab5bd81d8f6f9a71b1e68acf034c38b64d05878"}, "3": {"node_id": "89b8dfba-f448-47bf-9f08-711337807584", "node_type": "1", "metadata": {"page_label": "487", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "0e8db2dc7df6f48c5a0ee2855630bb70215954205d023bfc5573cf245bc91165"}}, "hash": "3331eb2e9de376bd019f697988de00aee747d7f51314e26edfcb48f88996e1f2", "text": "From Gaps to Transformation Paths in Enterprise Architecture Planning 487\nfor selecting actions is to allow an architect to select certain actions that result\nin a target. Furthermore, the tool is capable to provide metrics for quantitativeanalysis of the application landscape.\nGringel and Postina state that gap analysis needs a \u201cdetailed level of descrip-\ntion when it comes to modeling both landscapes\u201d ([ 27], p. 283) and as a result\nthe \u201cdata necessary to perform gap analysis on the entire application landscape\non a detailed level considering operations is overwhelming\u201d ([ 27], p. 291). How\nthe di\ufb00erent actions interfere with each other is not considered and actions canonly be provided if an ideal landscape with all details has been modeled.\n7.2 Strategic IT Managment by Hanschke\nThe \u2018Strategic IT Management\u2019 [ 11] approach is intended to serve as a toolkit\nfor EAM by providing best-practices derived from work experience. After a tar-\nget architecture has been modeled and agreed upon gap analysis is used to\ndetect di\ufb00erences between the current and target architecture. Gap analysis isperformed on the basis of process support maps visualizing which applications\nsupport which business processes (x-axis) and which customer group (y-axis)\nthe applications are assigned to. For a more \ufb01ne grained gap analysis Hanschkesuggests to additionally add information about services and information objects\nof the applications. Afterwards, for each gap possible actions to close the gap\nare considered.\nThe actions range from introducing a new application, adding or reducing\nfunctionality of an existing application, changing or adding services to the shut\ndown of applications and services. Based upon the results of gap analysis and\nderivation of appropriate actions it is necessary to clarify dependencies between\nthe actions, bundle the actions and create planned architectures as recommenda-tions for change. As far as we were able to verify the limitations of the tool and\napproach it is not possible to create suggestions for a detailed target architecture.\n7.3 ArchiMate\nArchiMate ([ 21], Chap. 11) introduces an Implementation and Migration Exten-\nsion", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "89b8dfba-f448-47bf-9f08-711337807584": {"__data__": {"id_": "89b8dfba-f448-47bf-9f08-711337807584", "embedding": null, "metadata": {"page_label": "487", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of conducting a gap analysis in enterprise architecture planning?\n2. What are the limitations of using ArchiMate for creating detailed target architectures?\n3. How can actions be used to support the creation of transition architectures as plateaus between the current and target architecture?", "prev_section_summary": "The section discusses the limitations of using gap analysis for selecting actions in Enterprise Architecture Planning (EAP). Gap analysis requires a detailed level of description when modeling both landscapes, and the data necessary to perform gap analysis on the entire application landscape on a detailed level considering operations is overwhelming. The Strategic IT Management approach uses gap analysis to detect differences between the current and target architecture, and possible actions to close the gaps identified through gap analysis are considered. The actions range from introducing a new application, adding or reducing functionality of an existing application, changing or adding services to the shut down of applications and services. The section also mentions the ArchiMate Implementation and Migration Extension, which introduces new concepts and modeling capabilities for EAP.", "section_summary": "The section discusses the purpose of conducting a gap analysis in enterprise architecture planning, the limitations of using ArchiMate for creating detailed target architectures, and how actions can be used to support the creation of transition architectures as plateaus between the current and target architecture. The section also introduces the concept of a gap in ArchiMate and its role in linking elements of two EA models. The section concludes with a discussion of future work in identifying situations where actions are necessary to create transition architectures as plateaus.", "excerpt_keywords": "1. Gap analysis\n2. ArchiMate\n3. Implementation and Migration Extension\n4. Gap element\n5. Core elements\n6. EA models\n7. Transition architectures\n8. Plateaus\n9. Stable states\n10. Architectural change"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "672b007c-4b9d-4463-a5d6-c4e1be707ef7", "node_type": "4", "metadata": {"page_label": "487", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "06a251c9de1a6d9955e088165ab5bd81d8f6f9a71b1e68acf034c38b64d05878"}, "2": {"node_id": "117b0747-5d77-4e73-a31a-003d96ed1e86", "node_type": "1", "metadata": {"page_label": "487", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "3331eb2e9de376bd019f697988de00aee747d7f51314e26edfcb48f88996e1f2"}}, "hash": "0e8db2dc7df6f48c5a0ee2855630bb70215954205d023bfc5573cf245bc91165", "text": "application, adding or reducing\nfunctionality of an existing application, changing or adding services to the shut\ndown of applications and services. Based upon the results of gap analysis and\nderivation of appropriate actions it is necessary to clarify dependencies between\nthe actions, bundle the actions and create planned architectures as recommenda-tions for change. As far as we were able to verify the limitations of the tool and\napproach it is not possible to create suggestions for a detailed target architecture.\n7.3 ArchiMate\nArchiMate ([ 21], Chap. 11) introduces an Implementation and Migration Exten-\nsion including a Gap element. A gap can be associated with any core element\nof the ArchiMate meta-models, except for the Value and Meaning element. In\ngeneral, a gap links several elements of two EA models and contains elements to\nbe removed (retired) and to be added (developed). The linkage of the di\ufb00erences\nbetween the EA models and the resulting gap is not described.\n8 Future Work\nCreating transition architectures as plateaus (see [ 21]) between the current and\ntarget architecture should be supported by actions. A plateau is a stable state\nof the EA. The current and target architecture are also plateaus according to\nArchiMate. However, we need to identify at \ufb01rst in which situations actions are", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "a9fada8e-463e-4063-8ce5-6f47fc12f666": {"__data__": {"id_": "a9fada8e-463e-4063-8ce5-6f47fc12f666", "embedding": null, "metadata": {"page_label": "488", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the paper \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\" and how does it relate to transition architecture creation and domain expert support?\n2. How does the value-based weighting methodology in the paper support a domain expert in determining which transformation paths are more promising?\n3. What are the limitations and future work areas identified in the paper, and how do they relate to related work in the field of enterprise architecture planning?", "prev_section_summary": "The section discusses the purpose of conducting a gap analysis in enterprise architecture planning, the limitations of using ArchiMate for creating detailed target architectures, and how actions can be used to support the creation of transition architectures as plateaus between the current and target architecture. The section also introduces the concept of a gap in ArchiMate and its role in linking elements of two EA models. The section concludes with a discussion of future work in identifying situations where actions are necessary to create transition architectures as plateaus.", "section_summary": "The section discusses a paper titled \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\" and its relevance to transition architecture creation and domain expert support. The paper presents a value-based weighting methodology to rank transformation paths based on different factors relevant for transformation planning. The action repository is also extended to cope with different EA models and concerns. The authors show how it is possible to get from identified gaps to transformation paths by creating a transformation model, detailing a target architecture, and using an action repository to create possible sequences of actions for transformation paths. They present a use case for parts of an application architecture and discuss the results and limitations of the solution, as well as future work to be addressed. The section references related work in the field of enterprise architecture planning, including ISO/IEC 42010, Winter and Fischer's \"Essential layers, artifacts, and dependencies of enterprise architecture,\" and Pulkkinen and Hirvonen's \"EA planning, development, and management process for agile enterprise development.\"", "excerpt_keywords": "1. Transformation architecture\n2. Action repository\n3. Target architecture\n4. Enterprise architecture\n5. Software engineering\n6. Systems design\n7. Architectural description\n8. Technical realization\n9. Agile enterprise development\n10. Systemic management"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "ca0fa447-93f1-41a7-b071-603606dd300e", "node_type": "4", "metadata": {"page_label": "488", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "b0385e450f438088394742c5e283c6ae66d5d576d1d4abf3491473344c886b57"}, "3": {"node_id": "74c41a08-73c9-4d8e-8c3f-aa24daa056f7", "node_type": "1", "metadata": {"page_label": "488", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "ae590a26baf3041d9a0adc6974856f733276247b209267448325b3748a4c6bf0"}}, "hash": "0c49270fff53a75436702a517c26a4558e93d6dbc7591ab9d139f5b5e75743e9", "text": "488 P. Diefenthaler and B. Bauer\nof relevance for transition architecture creation and if it is possible to provide\nmeaningful support for a domain expert.\nA value based weighting for di\ufb00erent transformation paths is currently elab-\norated to support a domain expert with information which paths seem to be\nmore promising than others. This ranking will take into account di\ufb00erent factorsrelevant for transformation planning.\nThe methodology how to create, use and maintain the action repository is\ncurrently extended to cope with di\ufb00erent EA models and di\ufb00erent concerns whichneed to be addressed during transformation planning.\n9 Conclusions\nWe have shown how it is possible to get from identi\ufb01ed gaps to transformationpaths by creating a transformation model, detailing a target architecture and\nusing an action repository to create possible sequences of actions for transfor-\nmation paths.\nAn use case for parts of an application architecture was presented and the\nsolution was applied to it. Furthermore, we presented a proposition for a tech-\nnical realisation to allow for tool support.\nWe discussed the results and limitations of the solution and clari\ufb01ed its con-\nnection to related work. Future work to be addressed was also presented.\nReferences\n1. International Organization for Standardization. ISO/IEC 42010:2007 Standard for\nsystems and software engineering - recommended practice for architectural descrip-tion of software-intensive systems (2007)\n2. Winter, R., Fischer, R.: Essential layers, artifacts, and dependencies of enterprise\narchitecture. In: 2006 10th IEEE International Enterprise Distributed Object Com-\nputing Conference Workshops (EDOCW\u201906), IEEE, p. 30 (2006)\n3. Pulkkinen, M., Hirvonen, A.: Ea planning, development and management process\nfor agile enterprise development. In: Proceedings of the 38th Annual Hawaii Inter-\nnational Conference on System Sciences, IEEE, p. 223c (2005)\n4. Pulkkinen, M.: Systemic management of architectural decisions in enterprise archi-\ntecture planning, four dimensions and three abstraction levels. In:", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "74c41a08-73c9-4d8e-8c3f-aa24daa056f7": {"__data__": {"id_": "74c41a08-73c9-4d8e-8c3f-aa24daa056f7", "embedding": null, "metadata": {"page_label": "488", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the development and management process for agile enterprise development, according to Pulkkinen and Hirvonen?\n2. How does Pulkkinen propose to systematically manage architectural decisions in enterprise architecture planning, considering four dimensions and three abstraction levels?\n3. What are the complexity levels of representing dynamics in enterprise architecture planning, according to Aier, Gleichauf, and Saat?", "prev_section_summary": "The section discusses a paper titled \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\" and its relevance to transition architecture creation and domain expert support. The paper presents a value-based weighting methodology to rank transformation paths based on different factors relevant for transformation planning. The action repository is also extended to cope with different EA models and concerns. The authors show how it is possible to get from identified gaps to transformation paths by creating a transformation model, detailing a target architecture, and using an action repository to create possible sequences of actions for transformation paths. They present a use case for parts of an application architecture and discuss the results and limitations of the solution, as well as future work to be addressed. The section references related work in the field of enterprise architecture planning, including ISO/IEC 42010, Winter and Fischer's \"Essential layers, artifacts, and dependencies of enterprise architecture,\" and Pulkkinen and Hirvonen's \"EA planning, development, and management process for agile enterprise development.\"", "section_summary": "The section discusses various topics related to enterprise architecture planning, including agile enterprise development, management of architectural decisions, complexity levels of representing dynamics, and IT governance. The authors and their works mentioned in the section include Pulkkinen, Hirvonen, Niemann, Aier, Gleichauf, Saat, and Winter. The section also mentions conferences and workshops where these authors presented their work, such as EDOCW'06 and HICSS'06.", "excerpt_keywords": "Enterprise architecture, Agile development, Systemic management, Architectural decisions, EA planning, Abstraction levels, IT governance, Effective IT management, Complexity levels, Dynamic transformation."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "ca0fa447-93f1-41a7-b071-603606dd300e", "node_type": "4", "metadata": {"page_label": "488", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "b0385e450f438088394742c5e283c6ae66d5d576d1d4abf3491473344c886b57"}, "2": {"node_id": "a9fada8e-463e-4063-8ce5-6f47fc12f666", "node_type": "1", "metadata": {"page_label": "488", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "0c49270fff53a75436702a517c26a4558e93d6dbc7591ab9d139f5b5e75743e9"}}, "hash": "ae590a26baf3041d9a0adc6974856f733276247b209267448325b3748a4c6bf0", "text": "In: 2006 10th IEEE International Enterprise Distributed Object Com-\nputing Conference Workshops (EDOCW\u201906), IEEE, p. 30 (2006)\n3. Pulkkinen, M., Hirvonen, A.: Ea planning, development and management process\nfor agile enterprise development. In: Proceedings of the 38th Annual Hawaii Inter-\nnational Conference on System Sciences, IEEE, p. 223c (2005)\n4. Pulkkinen, M.: Systemic management of architectural decisions in enterprise archi-\ntecture planning, four dimensions and three abstraction levels. In: Proceedings ofthe 39th Annual Hawaii International Conference on System Sciences (HICSS\u201906),\nIEEE, p. 179a (2006)\n5. Niemann, K.D.: From Enterprise Architecture to IT Governance: Elements of E\ufb00ec-\ntive IT Management. Vieweg, Wiesbaden (2006)\n6. Aier, S., Gleichauf, B., Saat, J., Winter, R.: Complexity levels of representing\ndynamics in EA planning. In: Albani, A., Barjis, J., Dietz, J.L.G. (eds.) Advances\nin Enterprise Engineering III. LNBIP, vol. 34, pp. 55\u201369. Springer, Heidelberg\n(2009)\n7. Aier, S., Gleichauf, B.: Towards a systematic approach for capturing dynamic\ntransformation in enterprise models. In: Sprague, R.H. (ed.) Proceedings of the\n43rd Hawaii International Conference on System Sciences 2010 (HICSS-43). Los\nAlamitos, IEEE Computer Society (2010)", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "cbf994fe-1a71-4fea-bf84-67df6b4bab6e": {"__data__": {"id_": "cbf994fe-1a71-4fea-bf84-67df6b4bab6e", "embedding": null, "metadata": {"page_label": "489", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the state-of-the-art in enterprise architecture management literature, and how has it evolved over time?\n2. How can enterprise models be applied to facilitate enterprise transformation, and what are some examples of their application?\n3. What is the role of information models in capturing the managed evolution of application landscapes, and how have they been used in this context?", "prev_section_summary": "The section discusses various topics related to enterprise architecture planning, including agile enterprise development, management of architectural decisions, complexity levels of representing dynamics, and IT governance. The authors and their works mentioned in the section include Pulkkinen, Hirvonen, Niemann, Aier, Gleichauf, Saat, and Winter. The section also mentions conferences and workshops where these authors presented their work, such as EDOCW'06 and HICSS'06.", "section_summary": "The section discusses the state-of-the-art in enterprise architecture management literature and its evolution over time. It also explores the application of enterprise models in facilitating enterprise transformation, and the role of information models in capturing the managed evolution of application landscapes. The section also mentions specific examples and tools used in these contexts, such as OWL web ontology language, SPARQL query language, and RDF Primer. Additionally, the section touches on semantic business process modeling and embedded EL+reasoning on programmable logic controllers.", "excerpt_keywords": "Enterprise architecture, management, transformation, models, information models, application landscapes, evolution, strategic management, IT, semantic web, web ontology, SPARQL, RDF, business process modeling, embedded reasoning, programmable logic controllers."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "2463fbe5-c01e-4300-b6ec-eca629caa29b", "node_type": "4", "metadata": {"page_label": "489", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "710cb5ff4959b9c99a37e3173bbab0cd10febb03f71cae910fd6e6acecc99a53"}, "3": {"node_id": "14415354-caa9-443b-9f88-a0e545afff92", "node_type": "1", "metadata": {"page_label": "489", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "e0265a62ff9d865824af917e710f277a3c9e6e446605ae199b1b617bf32c7e1f"}}, "hash": "38e092916bc66c29ddc919918fa10b4570bc96cc32a898ab0026d504140b8430", "text": "From Gaps to Transformation Paths in Enterprise Architecture Planning 489\n8. Buckl, S., Schweda, C.M.: On the State-of-the-Art in Enterprise Architecture Man-\nagement Literature (2011)\n9. Aier, S., Gleichauf, B.: Application of enterprise models for engineering enterprise\ntransformation. Enterp. Model. Inf. Syst. Archit. 5, 56\u201372 (2010)\n10. Buckl, S., Ernst, A.M., Matthes, F., Schweda, C.M.: An information model captur-\ning the managed evolution of application landscapes. J. Enterp. Archit. 5, 12\u201326\n(2009)\n11. Hanschke, I.: Strategisches Management der IT-Landschaft: Ein praktischer Leit-\nfaden f\u00a8 ur das Enterprise Architecture Management, 1st edn. Hanser, M\u00a8 unchen\n(2009)\n12. Shadbolt, N., Hall, W., Berners-Lee, T.: The semantic web revisited. IEEE Intell.\nSyst. 21(3), 96\u2013101 (2006). (IEEE Computer Society)\n13. Motik, B., Patel-Schneider, P.F., Horrocks, I.: Owl 2 web ontology language: Struc-\ntural speci\ufb01cation and functional-style syntax (2009)\n14. Prud\u2019hommeaux, E., Seaborne, A.: SPARQL Query Language for RDF. World\nWide Web Consortium (2008)\n15. Manola, F., Miller, E., McBride, B.: RDF Primer. World Wide Web Consortium\n(2004)\n16. Lautenbacher, F.: Semantic Business Process Modeling: Principles, Design Support\nand Realization. Shaker, Aachen (2010)\n17. Grimm, S., Watzke, M., Hubauer, T., Cescolini, F.: Embedded EL+reasoning\non programmable logic controllers. In: Cudr\u00b4 e-Mauroux, P.,", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "14415354-caa9-443b-9f88-a0e545afff92": {"__data__": {"id_": "14415354-caa9-443b-9f88-a0e545afff92", "embedding": null, "metadata": {"page_label": "489", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the relationship between RDF, Semantic Business Process Modeling, and Graph Neural Networks in the context of corporate relative valuation?\n2. How does the use of programmable logic controllers in embedded EL+reasoning relate to the field of Artificial Intelligence and its applications in business processes?\n3. What are the key differences between TOGAF and Archimate in terms of their specifications for Enterprise Architecture Planning and how do they relate to the field of Graph Transformation?", "prev_section_summary": "The section discusses the state-of-the-art in enterprise architecture management literature and its evolution over time. It also explores the application of enterprise models in facilitating enterprise transformation, and the role of information models in capturing the managed evolution of application landscapes. The section also mentions specific examples and tools used in these contexts, such as OWL web ontology language, SPARQL query language, and RDF Primer. Additionally, the section touches on semantic business process modeling and embedded EL+reasoning on programmable logic controllers.", "section_summary": "The section discusses the relationship between RDF, Semantic Business Process Modeling, and Graph Neural Networks in the context of corporate relative valuation. It also explores the use of programmable logic controllers in embedded EL+reasoning and its relation to Artificial Intelligence and its applications in business processes. The section compares and contrasts TOGAF and Archimate in terms of their specifications for Enterprise Architecture Planning and their relation to the field of Graph Transformation. Finally, the section discusses a survey of Enterprise Architecture Management Tools.", "excerpt_keywords": "RDF, Semantic Business Process Modeling, Embedded EL+reasoning, Artificial Intelligence, Automated Planning, TOGAF, Archimate, Enterprise Architecture Management Tool Survey, Graph transformation, Semantic Web."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "2463fbe5-c01e-4300-b6ec-eca629caa29b", "node_type": "4", "metadata": {"page_label": "489", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "710cb5ff4959b9c99a37e3173bbab0cd10febb03f71cae910fd6e6acecc99a53"}, "2": {"node_id": "cbf994fe-1a71-4fea-bf84-67df6b4bab6e", "node_type": "1", "metadata": {"page_label": "489", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "38e092916bc66c29ddc919918fa10b4570bc96cc32a898ab0026d504140b8430"}, "3": {"node_id": "4c21dca7-54f7-4706-a32a-350f74581f7e", "node_type": "1", "metadata": {"page_label": "489", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "df13aa9016434e6003c7d97406f6183174a1b7eeb740d92837f2dd25b477c632"}}, "hash": "e0265a62ff9d865824af917e710f277a3c9e6e446605ae199b1b617bf32c7e1f", "text": "RDF. World\nWide Web Consortium (2008)\n15. Manola, F., Miller, E., McBride, B.: RDF Primer. World Wide Web Consortium\n(2004)\n16. Lautenbacher, F.: Semantic Business Process Modeling: Principles, Design Support\nand Realization. Shaker, Aachen (2010)\n17. Grimm, S., Watzke, M., Hubauer, T., Cescolini, F.: Embedded EL+reasoning\non programmable logic controllers. In: Cudr\u00b4 e-Mauroux, P., He\ufb02in, J., Sirin, E.,\nTudorache, T., Euzenat, J., Hauswirth, M., Parreira, J.X., Hendler, J., Schreiber,G., Bernstein, A., Blomqvist, E. (eds.) ISWC 2012, Part II. LNCS, vol. 7650, pp.\n66\u201381. Springer, Heidelberg (2012)\n18. Russell, S.J., Norvig, P.: Arti\ufb01cial Intelligence: A Modern Approach, 3rd edn. Pren-\ntice Hall, Upper Saddle River (2010)\n19. Ghallab, M., Nau, D.S., Traverso, P.: Automated Planning: Theory & Practice.\nMorgan Kaufmann/Elsevier Science, San Francisco/Oxford (2004)\n20. The Open Group: TOGAF Version 9.1. TOGAF Series, 1st edn. Van Haren Pub-\nlishing, Zaltbommel (2011)\n21. The Open Group: Archimate 2.0 Speci\ufb01cation. Van Haren Publishing, Zaltbommel\n(2012)\n22. Matthes, F., Buckl, S., Leitel, J., Schweda, C.M.: Enterprise Architecture Manage-\nment Tool Survey 2008. Technische Universit\u00a8 at M\u00a8unchen, M\u00a8 unchen (2008)\n23. Edelkamp, S., Rensink, A.: Graph transformation and", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "4c21dca7-54f7-4706-a32a-350f74581f7e": {"__data__": {"id_": "4c21dca7-54f7-4706-a32a-350f74581f7e", "embedding": null, "metadata": {"page_label": "489", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the relationship between graph transformation and AI planning in enterprise architecture management?\n2. How can enterprise topologies be improved through segmentation, graph transformation, and analysis strategies?\n3. What is the role of graph grammars and computing by graph transformation in enterprise architecture management?", "prev_section_summary": "The section discusses the relationship between RDF, Semantic Business Process Modeling, and Graph Neural Networks in the context of corporate relative valuation. It also explores the use of programmable logic controllers in embedded EL+reasoning and its relation to Artificial Intelligence and its applications in business processes. The section compares and contrasts TOGAF and Archimate in terms of their specifications for Enterprise Architecture Planning and their relation to the field of Graph Transformation. Finally, the section discusses a survey of Enterprise Architecture Management Tools.", "section_summary": "The section discusses the relationship between graph transformation and AI planning in enterprise architecture management. It also explores how enterprise topologies can be improved through segmentation, graph transformation, and analysis strategies. The role of graph grammars and computing by graph transformation in enterprise architecture management is also highlighted. The section includes references to various publications and research papers on the topic.", "excerpt_keywords": "enterprise architecture, graph transformation, knowledge engineering, AI planning, enterprise topologies, segmentation, analysis strategies, master data management, I-pattern, gap analysis, software engineering, knowledge representation, knowledge discovery, knowledge engineering methodologies, knowledge engineering techniques, knowledge engineering tools, knowledge engineering frameworks, knowledge engineering applications, knowledge engineering research, knowledge engineering practice, knowledge engineering principles, knowledge engineering methodologies and techniques."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "2463fbe5-c01e-4300-b6ec-eca629caa29b", "node_type": "4", "metadata": {"page_label": "489", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "710cb5ff4959b9c99a37e3173bbab0cd10febb03f71cae910fd6e6acecc99a53"}, "2": {"node_id": "14415354-caa9-443b-9f88-a0e545afff92", "node_type": "1", "metadata": {"page_label": "489", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "e0265a62ff9d865824af917e710f277a3c9e6e446605ae199b1b617bf32c7e1f"}}, "hash": "df13aa9016434e6003c7d97406f6183174a1b7eeb740d92837f2dd25b477c632", "text": "Series, 1st edn. Van Haren Pub-\nlishing, Zaltbommel (2011)\n21. The Open Group: Archimate 2.0 Speci\ufb01cation. Van Haren Publishing, Zaltbommel\n(2012)\n22. Matthes, F., Buckl, S., Leitel, J., Schweda, C.M.: Enterprise Architecture Manage-\nment Tool Survey 2008. Technische Universit\u00a8 at M\u00a8unchen, M\u00a8 unchen (2008)\n23. Edelkamp, S., Rensink, A.: Graph transformation and AI Planning. In: Edelkamp,\nS., Frank, J. (eds.) Knowledge Engineering Competition (ICKEPS), Rhode Island,\nUSA (2007)\n24. Rozenberg, G.: Handbook of Graph Grammars and Computing by Graph Trans-\nformation, vol. 1. World Scienti\ufb01c River Edge, NJ, USA (1997)\n25. Binz, T., Leymann, F., Nowak, A., Schumm, D.: Improving the manageability of\nenterprise topologies through segmentation, graph transformation, and analysis\nstrategies. In: 2012 16th IEEE International Enterprise Distributed Object Com-puting Conference (EDOC 2012), pp. 61\u201370 (2012)\n26. Loshin, D.: Master Data Management. Elsevier/Morgan Kaufmann, Amster-\ndam/Boston (2009)\n27. Gringel, P., Postina, M.: I-pattern for gap analysis. In: Engels, G., Luckey, M.,\nPretschner, A., Reussner, R. (eds.) Software Engineering 2010. Lecture Notes in\nInformatics, pp. 281\u2013292. Gesellschaft f\u00a8 ur Informatik, Bonn (2010)", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "6f0b9627-227c-4eb5-92d1-4a57e4af6736": {"__data__": {"id_": "6f0b9627-227c-4eb5-92d1-4a57e4af6736", "embedding": null, "metadata": {"page_label": "1", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the Heterogeneous Graph Transformer (HGT) architecture and how does it differ from other graph neural network (GNN) architectures?\n2. How does the HGT architecture handle dynamic heterogeneous graphs and Web-scale graph data?\n3. What is the performance of the HGT model compared to other state-of-the-art GNN baselines on various downstream tasks?", "prev_section_summary": "The section discusses the relationship between graph transformation and AI planning in enterprise architecture management. It also explores how enterprise topologies can be improved through segmentation, graph transformation, and analysis strategies. The role of graph grammars and computing by graph transformation in enterprise architecture management is also highlighted. The section includes references to various publications and research papers on the topic.", "section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) architecture, which is designed for modeling heterogeneous graphs. The architecture is different from other graph neural network (GNN) architectures in that it handles heterogeneity by designing node- and edge-type dependent parameters. The HGT architecture also includes the relative temporal encoding technique to capture dynamic structural dependency and the heterogeneous mini-batch graph sampling algorithm (HGSampling) for efficient and scalable training. The section presents experimental results showing that the proposed HGT model outperforms state-of-the-art GNN baselines on various downstream tasks. The dataset and source code of HGT are publicly available. Key entities discussed in the section include graph neural networks, heterogeneous information networks, representation learning, graph embedding, and graph attention.", "excerpt_keywords": "Graph Neural Networks, Heterogeneous Information Networks, Representation Learning, Graph Embedding, Graph Attention, Web-scale Graph Data, Dynamic Heterogeneous Graphs, Relative Temporal Encoding, Heterogeneous Mini-batch Graph Sampling, Open Academic Graph."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "619d0736-2192-4da9-828a-f33ef9a047db", "node_type": "4", "metadata": {"page_label": "1", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "53ef4c20e509492cc140d37adee597ff6cd68aed6522a0a503af2eee93b5d6db"}, "3": {"node_id": "bea362d2-325d-497a-97d2-019c0ffa3e61", "node_type": "1", "metadata": {"page_label": "1", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "582d8920bde34c9914d665e629084c8a2324dbe721ac5b880a30156637330760"}}, "hash": "8544dc949beb6f03d6b8aa1300ec37be12b7e4e62a97302c5aeffcb8f1c3a074", "text": "Heterogeneous Graph Transformer\nZiniu Hu\u2217\nUniversity of California, Los Angeles\nbull@cs.ucla.eduYuxiao Dong\nMicrosoft Research, Redmond\nyuxdong@microsoft.com\nKuansan Wang\nMicrosoft Research, Redmond\nkuansanw@microsoft.comYizhou Sun\nUniversity of California, Los Angeles\nyzsun@cs.ucla.edu\nABSTRACT\nRecent years have witnessed the emerging success of graph neu-\nral networks (GNNs) for modeling structured data. However, most\nGNNs are designed for homogeneous graphs, in which all nodes\nand edges belong to the same types, making them infeasible to\nrepresent heterogeneous structures. In this paper, we present the\nHeterogeneous Graph Transformer (HGT) architecture for mod-\neling Web-scale heterogeneous graphs. To model heterogeneity,\nwe design node- and edge-type dependent parameters to charac-\nterize the heterogeneous attention over each edge, empowering\nHGT to maintain dedicated representations for different types of\nnodes and edges. To handle dynamic heterogeneous graphs, we in-\ntroduce the relative temporal encoding technique into HGT, which\nis able to capture the dynamic structural dependency with arbitrary\ndurations. To handle Web-scale graph data, we design the hetero-\ngeneous mini-batch graph sampling algorithm\u2014HGSampling\u2014for\nefficient and scalable training. Extensive experiments on the Open\nAcademic Graph of 179 million nodes and 2 billion edges show\nthat the proposed HGT model consistently outperforms all the\nstate-of-the-art GNN baselines by 9 %\u201321%on various downstream\ntasks. The dataset and source code of HGT are publicly available at\nhttps://github.com/acbull/pyHGT.\nKEYWORDS\nGraph Neural Networks; Heterogeneous Information Networks;\nRepresentation Learning; Graph Embedding; Graph Attention\nACM Reference Format:\nZiniu Hu, Yuxiao Dong, Kuansan Wang, and Yizhou Sun. 2020. Hetero-\ngeneous Graph Transformer. In Proceedings of The Web Conference 2020\n(WWW \u201920),", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "bea362d2-325d-497a-97d2-019c0ffa3e61": {"__data__": {"id_": "bea362d2-325d-497a-97d2-019c0ffa3e61", "embedding": null, "metadata": {"page_label": "1", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Graph Transformer (HGT) and how does it differ from other Graph Neural Networks (GNNs)?\n2. What are the benefits of using HGT for modeling complex systems with heterogeneous information networks?\n3. How does HGT handle the challenge of dealing with heterogeneous graphs that have objects of different types?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) architecture, which is designed for modeling heterogeneous graphs. The architecture is different from other graph neural network (GNN) architectures in that it handles heterogeneity by designing node- and edge-type dependent parameters. The HGT architecture also includes the relative temporal encoding technique to capture dynamic structural dependency and the heterogeneous mini-batch graph sampling algorithm (HGSampling) for efficient and scalable training. The section presents experimental results showing that the proposed HGT model outperforms state-of-the-art GNN baselines on various downstream tasks. The dataset and source code of HGT are publicly available. Key entities discussed in the section include graph neural networks, heterogeneous information networks, representation learning, graph embedding, and graph attention.", "section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT), a type of Graph Neural Network (GNN) that is used to model complex systems with heterogeneous information networks. HGT differs from other GNNs in that it takes only the one-hop edges of a heterogeneous graph as input and does not require manual design of meta paths. The benefits of using HGT for modeling complex systems include improved performance on various downstream tasks and the ability to handle heterogeneous graphs with objects of different types. The dataset and source code for HGT are publicly available at <https://github.com/acbull/pyHGT>. The section also discusses the challenges of dealing with heterogeneous graphs and how HGT addresses them.", "excerpt_keywords": "Graph Neural Networks, Heterogeneous Information Networks, Representation Learning, Graph Embedding, Graph Attention, Web-scale, Academic Network, Meta Paths, Meta Relations, One-hop Edges, Abstracting, Modeling, Complex Systems."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "619d0736-2192-4da9-828a-f33ef9a047db", "node_type": "4", "metadata": {"page_label": "1", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "53ef4c20e509492cc140d37adee597ff6cd68aed6522a0a503af2eee93b5d6db"}, "2": {"node_id": "6f0b9627-227c-4eb5-92d1-4a57e4af6736", "node_type": "1", "metadata": {"page_label": "1", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "8544dc949beb6f03d6b8aa1300ec37be12b7e4e62a97302c5aeffcb8f1c3a074"}, "3": {"node_id": "13b2c56c-d4b4-4173-b755-8b9800d99c3d", "node_type": "1", "metadata": {"page_label": "1", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "e6cfce1adf21357882dd73f52514f3b9cca9fc9e63e54b5bbc9b62e656bfcea7"}}, "hash": "582d8920bde34c9914d665e629084c8a2324dbe721ac5b880a30156637330760", "text": "GNN baselines by 9 %\u201321%on various downstream\ntasks. The dataset and source code of HGT are publicly available at\nhttps://github.com/acbull/pyHGT.\nKEYWORDS\nGraph Neural Networks; Heterogeneous Information Networks;\nRepresentation Learning; Graph Embedding; Graph Attention\nACM Reference Format:\nZiniu Hu, Yuxiao Dong, Kuansan Wang, and Yizhou Sun. 2020. Hetero-\ngeneous Graph Transformer. In Proceedings of The Web Conference 2020\n(WWW \u201920), April 20\u201324, 2020, Taipei, Taiwan. ACM, New York, NY, USA,\n11 pages. https://doi.org/10.1145/3366423.3380027\n1 INTRODUCTION\nHeterogeneous graphs have been commonly used for abstracting\nand modeling complex systems, in which objects of different types\n\u2217This work was done when Ziniu was an intern at Microsoft Research.\nPermission to make digital or hard copies of all or part of this work for personal or\nclassroom use is granted without fee provided that copies are not made or distributed\nfor profit or commercial advantage and that copies bear this notice and the full citation\non the first page. Copyrights for components of this work owned by others than ACM\nmust be honored. Abstracting with credit is permitted. To copy otherwise, or republish,\nto post on servers or to redistribute to lists, requires prior specific permission and/or a\nfee. Request permissions from permissions@acm.org.\nWWW \u201920, April 20\u201324, 2020, Taipei, Taiwan\n\u00a92020 Association for Computing Machinery.\nACM ISBN 978-1-4503-7023-3/20/04.\nhttps://doi.org/10.1145/3366423.3380027\nFigure 1: The schema and meta relations of Open Academic\nGraph (OAG). Given a Web-scale heterogeneous graph, e.g., an\nacademic network, HGT takes only its one-hop edges as input\nwithout manually designing meta paths.\ninteract with each other in various ways. Some prevalent", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "13b2c56c-d4b4-4173-b755-8b9800d99c3d": {"__data__": {"id_": "13b2c56c-d4b4-4173-b755-8b9800d99c3d", "embedding": null, "metadata": {"page_label": "1", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Graph Transformer (HGT) and how does it differ from other approaches to mining heterogeneous graphs?\n2. What are some common types of nodes and relationships found in heterogeneous graphs, and how do they interact with each other?\n3. What are some of the challenges faced by existing approaches to mining heterogeneous graphs, and how does HGT address these challenges?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT), a type of Graph Neural Network (GNN) that is used to model complex systems with heterogeneous information networks. HGT differs from other GNNs in that it takes only the one-hop edges of a heterogeneous graph as input and does not require manual design of meta paths. The benefits of using HGT for modeling complex systems include improved performance on various downstream tasks and the ability to handle heterogeneous graphs with objects of different types. The dataset and source code for HGT are publicly available at <https://github.com/acbull/pyHGT>. The section also discusses the challenges of dealing with heterogeneous graphs and how HGT addresses them.", "section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) and its differences from other approaches to mining heterogeneous graphs. It also describes the types of nodes and relationships found in heterogeneous graphs and the challenges faced by existing approaches to mining them. The section highlights the Open Academic Graph (OAG) as an example of a heterogeneous graph and explains how HGT addresses the challenges faced by existing approaches.", "excerpt_keywords": "academic graph, heterogeneous graph, graph neural networks, meta paths, PathSim, metapath2vec, Open Academic Graph, nodes, edges, dynamic nature, Web-scale heterogeneous graphs, domain knowledge, feature and representation space, non-sharing weights, intrinsic design, implementation."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "619d0736-2192-4da9-828a-f33ef9a047db", "node_type": "4", "metadata": {"page_label": "1", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "53ef4c20e509492cc140d37adee597ff6cd68aed6522a0a503af2eee93b5d6db"}, "2": {"node_id": "bea362d2-325d-497a-97d2-019c0ffa3e61", "node_type": "1", "metadata": {"page_label": "1", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "582d8920bde34c9914d665e629084c8a2324dbe721ac5b880a30156637330760"}, "3": {"node_id": "84388c23-78e7-4d73-9103-a443e9587e62", "node_type": "1", "metadata": {"page_label": "1", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "0af21273726309b0bd66dab07087b2d92fe340084e862ed967ebdad337345328"}}, "hash": "e6cfce1adf21357882dd73f52514f3b9cca9fc9e63e54b5bbc9b62e656bfcea7", "text": "April 20\u201324, 2020, Taipei, Taiwan\n\u00a92020 Association for Computing Machinery.\nACM ISBN 978-1-4503-7023-3/20/04.\nhttps://doi.org/10.1145/3366423.3380027\nFigure 1: The schema and meta relations of Open Academic\nGraph (OAG). Given a Web-scale heterogeneous graph, e.g., an\nacademic network, HGT takes only its one-hop edges as input\nwithout manually designing meta paths.\ninteract with each other in various ways. Some prevalent instances\nof such systems include academic graphs, Facebook entity graph,\nLinkedIn economic graph, and broadly the Internet of Things net-\nwork. For example, the Open Academic Graph (OAG) [ 28] in Figure\n1 contains five types of nodes: papers, authors, institutions, venues\n(journal, conference, or preprint), and fields, as well as different\ntypes of relationships between them.\nOver the past decade, a significant line of research has been ex-\nplored for mining heterogeneous graphs [ 17]. One of the classical\nparadigms is to define and use meta paths to model heterogeneous\nstructures, such as PathSim [ 18] and metapath2vec [ 3]. Recently,\nin view of graph neural networks\u2019 (GNNs) success [ 7,9,22], there\nare several attempts to adopt GNNs to learn with heterogeneous\nnetworks [ 14,23,26,27]. However, these works face several issues:\nFirst, most of them involve the design of meta paths for each type of\nheterogeneous graphs, requiring specific domain knowledge; Sec-\nond, they either simply assume that different types of nodes/edges\nshare the same feature and representation space or keep distinct\nnon-sharing weights for either node type or edge type alone, mak-\ning them insufficient to capture heterogeneous graphs\u2019 properties;\nThird, most of them ignore the dynamic nature of every (hetero-\ngeneous) graph; Finally, their intrinsic design and implementation\nmake them incapable of modeling Web-scale heterogeneous graphs.\nTake OAG for example: First, the nodes and edges in", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "84388c23-78e7-4d73-9103-a443e9587e62": {"__data__": {"id_": "84388c23-78e7-4d73-9103-a443e9587e62", "embedding": null, "metadata": {"page_label": "1", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What are the limitations of existing heterogeneous graph neural network models, and how does the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation address these limitations?\n2. How does the dynamic nature of heterogeneous graphs affect the design and implementation of graph neural network models, and how does the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation account for this dynamic nature?\n3. Can you provide an example of a heterogeneous graph, such as OAG, and explain how the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation can capture the unique properties of this graph that other models may struggle with?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) and its differences from other approaches to mining heterogeneous graphs. It also describes the types of nodes and relationships found in heterogeneous graphs and the challenges faced by existing approaches to mining them. The section highlights the Open Academic Graph (OAG) as an example of a heterogeneous graph and explains how HGT addresses the challenges faced by existing approaches.", "section_summary": "The section discusses the limitations of existing heterogeneous graph neural network models and introduces the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation. The model addresses the limitations by accounting for the dynamic nature of heterogeneous graphs, capturing unique properties of each type of heterogeneous graph, and modeling Web-scale heterogeneous graphs. The section uses the OAG as an example to illustrate the limitations of existing models and how the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation can capture its unique properties.", "excerpt_keywords": "heterogeneous graphs, domain knowledge, node/edge sharing, representation space, non-sharing weights, dynamic nature, Web-scale, OAG, nodes, edges, evolution, publications, KDD conference, machine learning, text features, affiliated scholars, citation links, volume, doubling, every 12 years, related, database, recent years."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "619d0736-2192-4da9-828a-f33ef9a047db", "node_type": "4", "metadata": {"page_label": "1", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "53ef4c20e509492cc140d37adee597ff6cd68aed6522a0a503af2eee93b5d6db"}, "2": {"node_id": "13b2c56c-d4b4-4173-b755-8b9800d99c3d", "node_type": "1", "metadata": {"page_label": "1", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "e6cfce1adf21357882dd73f52514f3b9cca9fc9e63e54b5bbc9b62e656bfcea7"}}, "hash": "0af21273726309b0bd66dab07087b2d92fe340084e862ed967ebdad337345328", "text": "each type of\nheterogeneous graphs, requiring specific domain knowledge; Sec-\nond, they either simply assume that different types of nodes/edges\nshare the same feature and representation space or keep distinct\nnon-sharing weights for either node type or edge type alone, mak-\ning them insufficient to capture heterogeneous graphs\u2019 properties;\nThird, most of them ignore the dynamic nature of every (hetero-\ngeneous) graph; Finally, their intrinsic design and implementation\nmake them incapable of modeling Web-scale heterogeneous graphs.\nTake OAG for example: First, the nodes and edges in OAG could\nhave different feature distributions, e.g., papers have text features\nwhereas institutions may have features from affiliated scholars, and\ncoauthorships obviously differ from citation links; Second, OAG\nhas been consistently evolving, e.g., 1) the volume of publications\ndoubles every 12 years [ 4], and 2) the KDD conference was more\nrelated to database in the 1990s whereas more to machine learning\nin recent years; Finally, OAG contains hundreds of millions of nodesarXiv:2003.01332v1  [cs.LG]  3 Mar 2020", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "36c8ac35-d0df-4e32-888a-72bc36a252d2": {"__data__": {"id_": "36c8ac35-d0df-4e32-888a-72bc36a252d2", "embedding": null, "metadata": {"page_label": "2", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the main goal of the proposed Heterogeneous Graph Transformer (HGT) architecture?\n2. How does the node- and edge-type dependent attention mechanism in HGT handle graph heterogeneity?\n3. What is the relative temporal encoding (RTE) strategy proposed to handle graph dynamics in HGT?", "prev_section_summary": "The section discusses the limitations of existing heterogeneous graph neural network models and introduces the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation. The model addresses the limitations by accounting for the dynamic nature of heterogeneous graphs, capturing unique properties of each type of heterogeneous graph, and modeling Web-scale heterogeneous graphs. The section uses the OAG as an example to illustrate the limitations of existing models and how the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation can capture its unique properties.", "section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) architecture, which is proposed to handle the limitations and challenges of existing heterogeneous graph neural networks (GNNs) in dealing with large-scale heterogeneous graphs. The HGT architecture introduces a node- and edge-type dependent attention mechanism to maintain specific representation spaces for nodes and edges of different types, while allowing connected nodes in different types to interact and aggregate messages. The architecture also incorporates information from high-order neighbors of different types through message passing across layers, and proposes a relative temporal encoding (RTE) strategy to handle graph dynamics. The section highlights the benefits of the HGT architecture, including its ability to capture network dynamics, avoid customized meta paths, and be scalable to Web-scale graphs.", "excerpt_keywords": "1. Heterogeneous Graph Neural Networks\n2. Node- and Edge-type Dependent Representations\n3. Graph Dynamics\n4. Scalability\n5. Heterogeneous Mutual Attention\n6. Meta Relations\n7. Attention Mechanism\n8. Message Passing\n9. Relative Temporal Encoding\n10. Web-scale Graphs"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "a3e1e120-6732-4b4a-89ee-477130998e0e", "node_type": "4", "metadata": {"page_label": "2", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "0165fd3447625d17d4424f764da78f21b9aab377a01e12e0ae99a474bdcca254"}, "3": {"node_id": "e244e139-e9a1-4e34-92b2-5835ada25da6", "node_type": "1", "metadata": {"page_label": "2", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "65a1d54feff714ad767fd7186908f07f1a15df9d9fc4f9199a11093a0e6c722b"}}, "hash": "8138297f21fd97d93f848933c7cec8289ecb7c9358ffef0bbb52e3c74c0b8bff", "text": "WWW \u201920, April 20\u201324, 2020, Taipei, Taiwan Ziniu Hu, Yuxiao Dong, Kuansan Wang, and Yizhou Sun\nand billions of relationships, leaving existing heterogeneous GNNs\nnot scalable for handling it.\nIn light of these limitations and challenges, we propose to study\nheterogeneous graph neural networks with the goal of maintaining\nnode- and edge-type dependent representations, capturing network\ndynamics, avoiding customized meta paths, and being scalable to\nWeb-scale graphs. In this work, we present the Heterogeneous\nGraph Transformer (HGT) architecture to deal with all these issues.\nTo handle graph heterogeneity, we introduce the node- and edge-\ntype dependent attention mechanism. Instead of parameterizing\neach type of edges, the heterogeneous mutual attention in HGT is\ndefined by breaking down each edge e=(s,t)based on its meta\nrelation triplet, i.e., \u27e8node type of s, edge type of ebetween s&\nt, node type of t\u27e9. Figure 1 illustrates the meta relations of hetero-\ngeneous academic graphs. In specific, we use these meta relations\nto parameterize the weight matrices for calculating attention over\neach edge. As a result, nodes and edges of different types are al-\nlowed to maintain their specific representation spaces. Meanwhile,\nconnected nodes in different types can still interact, pass, and aggre-\ngate messages without being restricted by their distribution gaps.\nDue to the nature of its architecture, HGT can incorporate informa-\ntion from high-order neighbors of different types through message\npassing across layers, which can be regarded as \u201csoft\u201d meta paths.\nThat said, even if HGT take only its one-hop edges as input without\nmanually designing meta paths, the proposed attention mechanism\ncan automatically and implicitly learn and extract \u201cmeta paths\u201d that\nare important for different downstream tasks.\nTo handle graph dynamics, we enhance HGT by proposing the\nrelative temporal encoding (RTE) strategy. Instead of slicing the\ninput graph into different timestamps, we propose to maintain all\nthe edges happening in different times as a whole, and design", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "e244e139-e9a1-4e34-92b2-5835ada25da6": {"__data__": {"id_": "e244e139-e9a1-4e34-92b2-5835ada25da6", "embedding": null, "metadata": {"page_label": "2", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the proposed attention mechanism in Heterogeneous Graph Transformer (HGT) and how does it automatically learn and extract important meta paths for different downstream tasks?\n2. How does the relative temporal encoding (RTE) strategy in HGT handle graph dynamics and enable end-to-end learning of temporal dependencies and evolution of heterogeneous graphs?\n3. What is the Heterogeneous Sub-Graph Sampling (HGSampling) algorithm and how does it address the imbalance in node and edge types in existing GNN sampling methods for training and inference on arbitrary-size heterogeneous graphs?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) architecture, which is proposed to handle the limitations and challenges of existing heterogeneous graph neural networks (GNNs) in dealing with large-scale heterogeneous graphs. The HGT architecture introduces a node- and edge-type dependent attention mechanism to maintain specific representation spaces for nodes and edges of different types, while allowing connected nodes in different types to interact and aggregate messages. The architecture also incorporates information from high-order neighbors of different types through message passing across layers, and proposes a relative temporal encoding (RTE) strategy to handle graph dynamics. The section highlights the benefits of the HGT architecture, including its ability to capture network dynamics, avoid customized meta paths, and be scalable to Web-scale graphs.", "section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) model, which is a multi-modal graph neural network (GNN) for corporate relative valuation. The proposed attention mechanism in HGT automatically learns and extracts important meta paths for different downstream tasks. The relative temporal encoding (RTE) strategy in HGT handles graph dynamics and enables end-to-end learning of temporal dependencies and evolution of heterogeneous graphs. The Heterogeneous Sub-Graph Sampling (HGSampling) algorithm is designed to address the imbalance in node and edge types in existing GNN sampling methods for training and inference on arbitrary-size heterogeneous graphs. The effectiveness and efficiency of the proposed HGT model is demonstrated on the Web-scale Open Academic Graph and domain-specific graphs such as computer science and medicine academic graphs.", "excerpt_keywords": "1. Heterogeneous Graph Transformer, 2. Attention mechanism, 3. Relative temporal encoding, 4. Heterogeneous sub-graph sampling, 5. Web-scale graph data, 6. Node and edge types, 7. Representation learning, 8. Domain-specific graphs, 9. Computer science, 10. Medicine"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "a3e1e120-6732-4b4a-89ee-477130998e0e", "node_type": "4", "metadata": {"page_label": "2", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "0165fd3447625d17d4424f764da78f21b9aab377a01e12e0ae99a474bdcca254"}, "2": {"node_id": "36c8ac35-d0df-4e32-888a-72bc36a252d2", "node_type": "1", "metadata": {"page_label": "2", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "8138297f21fd97d93f848933c7cec8289ecb7c9358ffef0bbb52e3c74c0b8bff"}, "3": {"node_id": "bc9e8ddd-384f-4f71-b00b-acff631de8d2", "node_type": "1", "metadata": {"page_label": "2", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "543060de817214355f744990f57ac20fe01963d895658cd73bd8c906491a73ea"}}, "hash": "65a1d54feff714ad767fd7186908f07f1a15df9d9fc4f9199a11093a0e6c722b", "text": "types through message\npassing across layers, which can be regarded as \u201csoft\u201d meta paths.\nThat said, even if HGT take only its one-hop edges as input without\nmanually designing meta paths, the proposed attention mechanism\ncan automatically and implicitly learn and extract \u201cmeta paths\u201d that\nare important for different downstream tasks.\nTo handle graph dynamics, we enhance HGT by proposing the\nrelative temporal encoding (RTE) strategy. Instead of slicing the\ninput graph into different timestamps, we propose to maintain all\nthe edges happening in different times as a whole, and design the\nRTE strategy to model structural temporal dependencies with any\nduration length, and even with unseen and future timestamps. By\nend-to-end training, RTE enables HGT to automatically learn the\ntemporal dependency and evolution of heterogeneous graphs.\nTo handle Web-scale graph data, we design the first hetero-\ngeneous sub-graph sampling algorithm\u2014HGSampling\u2014for mini-\nbatch GNN training. Its main idea is to sample heterogeneous sub-\ngraphs in which different types of nodes are with similar propor-\ntions, since the direct usage of existing (homogeneous) GNN sam-\npling methods, such as GraphSage [ 7], FastGCN [ 1], and LADIES [ 29],\nresults in highly imbalanced ones regarding to both node and edge\ntypes. In addition, it is also designed to keep the sampled sub-graphs\ndense for minimizing the loss of information. With HGSampling,\nall the GNN models, including our proposed HGT, can train and\ninfer on arbitrary-size heterogeneous graphs.\nWe demonstrate the effectiveness and efficiency of the proposed\nHeterogeneous Graph Transformer on the Web-scale Open Aca-\ndemic Graph comprised of 179 million nodes and 2 billion edges\nspanning from 1900 to 2019, making this the largest-scale and\nlongest-spanning representation learning yet performed on hetero-\ngeneous graphs. Additionally, we also examine it on domain-specific\ngraphs: the computer science and medicine academic graphs. Exper-\nimental results suggest that HGT can significantly improve various\ndownstream tasks", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "bc9e8ddd-384f-4f71-b00b-acff631de8d2": {"__data__": {"id_": "bc9e8ddd-384f-4f71-b00b-acff631de8d2", "embedding": null, "metadata": {"page_label": "2", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the definition of a heterogeneous graph and how is it different from a homogeneous graph?\n2. What is the difference between Heterogeneous Graph Transformer (HGT) and existing heterogeneous graph neural networks?\n3. Can you explain the concept of meta relations and meta paths in the context of heterogeneous graphs?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) model, which is a multi-modal graph neural network (GNN) for corporate relative valuation. The proposed attention mechanism in HGT automatically learns and extracts important meta paths for different downstream tasks. The relative temporal encoding (RTE) strategy in HGT handles graph dynamics and enables end-to-end learning of temporal dependencies and evolution of heterogeneous graphs. The Heterogeneous Sub-Graph Sampling (HGSampling) algorithm is designed to address the imbalance in node and edge types in existing GNN sampling methods for training and inference on arbitrary-size heterogeneous graphs. The effectiveness and efficiency of the proposed HGT model is demonstrated on the Web-scale Open Academic Graph and domain-specific graphs such as computer science and medicine academic graphs.", "section_summary": "The section discusses the concept of heterogeneous graphs and their difference from homogeneous graphs. It also introduces the Heterogeneous Graph Transformer (HGT) and its effectiveness and efficiency in training and inferring on arbitrary-size heterogeneous graphs. The section reviews recent development on graph neural networks (GNNs) and their heterogeneous variants, and highlights the difference between HGT and existing attempts on heterogeneous graph neural networks. The section also explains the concept of meta relations and meta paths in the context of heterogeneous graphs.", "excerpt_keywords": "1. Heterogeneous Graphs, 2. Network Dynamics, 3. Graph Neural Networks (GNNs), 4. Heterogeneous Variants, 5. Heterogeneous Graph Mining, 6. Meta Relation, 7. Meta Path, 8. Complex Systems, 9. Real-World Data, 10. Representation Learning."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "a3e1e120-6732-4b4a-89ee-477130998e0e", "node_type": "4", "metadata": {"page_label": "2", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "0165fd3447625d17d4424f764da78f21b9aab377a01e12e0ae99a474bdcca254"}, "2": {"node_id": "e244e139-e9a1-4e34-92b2-5835ada25da6", "node_type": "1", "metadata": {"page_label": "2", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "65a1d54feff714ad767fd7186908f07f1a15df9d9fc4f9199a11093a0e6c722b"}, "3": {"node_id": "ade4f093-5318-40d7-9b81-2204d26fff8e", "node_type": "1", "metadata": {"page_label": "2", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "b270217354ba21822ccd1bdd1f5f30a1c001697cdd7e48ad448a3a6a590ebb9a"}}, "hash": "543060de817214355f744990f57ac20fe01963d895658cd73bd8c906491a73ea", "text": "can train and\ninfer on arbitrary-size heterogeneous graphs.\nWe demonstrate the effectiveness and efficiency of the proposed\nHeterogeneous Graph Transformer on the Web-scale Open Aca-\ndemic Graph comprised of 179 million nodes and 2 billion edges\nspanning from 1900 to 2019, making this the largest-scale and\nlongest-spanning representation learning yet performed on hetero-\ngeneous graphs. Additionally, we also examine it on domain-specific\ngraphs: the computer science and medicine academic graphs. Exper-\nimental results suggest that HGT can significantly improve various\ndownstream tasks over state-of-the-art GNNs as well as dedicated\nheterogeneous models by 9\u201321 %. We further conduct case studiesto show the proposed method can indeed automatically capture the\nimportance of implicit meta paths for different tasks.\n2 PRELIMINARIES AND RELATED WORK\nIn this section, we introduce the basic definition of heteroge-\nneous graphs with network dynamics and review the recent devel-\nopment on graph neural networks (GNNs) and their heterogeneous\nvariants. We also highlight the difference between HGT and existing\nattempts on heterogeneous graph neural networks.\n2.1 Heterogeneous Graph Mining\nHeterogeneous graphs [ 17] (a.k.a., heterogeneous information\nnetworks) are an important abstraction for modeling relational data\nfor many real-world complex systems. Formally, it is defined as:\nDefinition 1. Heterogeneous Graph: A heterogeneous graph\nis defined as a directed graph G=(V,E,A,R)where each node\nv\u2208V and each edge e\u2208Eare associated with their type mapping\nfunctions\u03c4(v):V\u2192A and\u03d5(e):E\u2192R , respectively.\nMeta Relation. For an edge e=(s,t)linked from source node sto\ntarget node t, its meta relation is denoted as \u27e8\u03c4(s),\u03d5(e),\u03c4(t)\u27e9. Natu-\nrally,\u03d5(e)\u22121represents the inverse of \u03d5(e). The classical meta path\nparadigm", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "ade4f093-5318-40d7-9b81-2204d26fff8e": {"__data__": {"id_": "ade4f093-5318-40d7-9b81-2204d26fff8e", "embedding": null, "metadata": {"page_label": "2", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. How does the heterogeneous graph transformer model handle multiple types of relations between different types of nodes in real-world networks?\n2. How does the dynamic heterogeneous graph model account for the timestamp of edges and the possibility of multiple timestamps for a node?\n3. Can you explain the difference between the first and upcoming editions of the \"WWW\" conference node in the dynamic heterogeneous graph model?", "prev_section_summary": "The section discusses the concept of heterogeneous graphs and their difference from homogeneous graphs. It also introduces the Heterogeneous Graph Transformer (HGT) and its effectiveness and efficiency in training and inferring on arbitrary-size heterogeneous graphs. The section reviews recent development on graph neural networks (GNNs) and their heterogeneous variants, and highlights the difference between HGT and existing attempts on heterogeneous graph neural networks. The section also explains the concept of meta relations and meta paths in the context of heterogeneous graphs.", "section_summary": "The section discusses the Heterogeneous Graph Transformer model, which is used to handle multiple types of relations between different types of nodes in real-world networks. The model accounts for the timestamp of edges and the possibility of multiple timestamps for a node through the dynamic heterogeneous graph model. The section also explains the difference between the first and upcoming editions of the \"WWW\" conference node in the dynamic heterogeneous graph model.", "excerpt_keywords": "heterogeneous networks, meta relation, dynamic heterogeneous graph, timestamp, multiple types of relations, authorship order, OAG, first author, second author, conference, WWW, internet protocol, Web infrastructure, social analysis, ubiquitous computing, search & IR, privacy and society, upcoming WWW."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "a3e1e120-6732-4b4a-89ee-477130998e0e", "node_type": "4", "metadata": {"page_label": "2", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "0165fd3447625d17d4424f764da78f21b9aab377a01e12e0ae99a474bdcca254"}, "2": {"node_id": "bc9e8ddd-384f-4f71-b00b-acff631de8d2", "node_type": "1", "metadata": {"page_label": "2", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "543060de817214355f744990f57ac20fe01963d895658cd73bd8c906491a73ea"}, "3": {"node_id": "8e1b103b-1d79-485b-8bbe-3161ba6db37f", "node_type": "1", "metadata": {"page_label": "2", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "7ec7bbbeae3a50b6d5b7e41c5fee182abb0d778ff4bbf8f6311b9b23e6d1fa12"}}, "hash": "b270217354ba21822ccd1bdd1f5f30a1c001697cdd7e48ad448a3a6a590ebb9a", "text": "each node\nv\u2208V and each edge e\u2208Eare associated with their type mapping\nfunctions\u03c4(v):V\u2192A and\u03d5(e):E\u2192R , respectively.\nMeta Relation. For an edge e=(s,t)linked from source node sto\ntarget node t, its meta relation is denoted as \u27e8\u03c4(s),\u03d5(e),\u03c4(t)\u27e9. Natu-\nrally,\u03d5(e)\u22121represents the inverse of \u03d5(e). The classical meta path\nparadigm [17\u201319] is defined as a sequence of such meta relation.\nNotice that, to better model real-world heterogeneous networks,\nwe assume that there may exist multiple types of relations between\ndifferent types of nodes. For example, in OAG there are different\ntypes of relations between the author andpaper nodes by consid-\nering the authorship order, i.e., \u201cthe first author of\u201d, \u201cthe second\nauthor of\u201d, and so on.\nDynamic Heterogeneous Graph. To model the dynamic nature\nof real-world (heterogeneous) graphs, we assign an edge e=(s,t)\na timestamp T, when node sconnects to node tatT. Ifsappears\nfor the first time, Tis also assigned to s.scan be associated with\nmultiple timestamps if it builds connections over time.\nIn other words, we assume that the timestamp of an edge is\nunchanged, denoting the time it is created. For example, when a\npaper published on a conference at time T,Twill be assigned to\nthe edge between the paper and conference nodes. On the contrary,\ndifferent timestamps can be assigned to a node accordingly. For\nexample, the conference node \u201cWWW\u201d can be assigned any year.\nWWW @1994 means that we are considering the first edition of\nWWW, which focuses more on internet protocol and Web infras-\ntructure, while WWW @2020 means the upcoming WWW, which\nexpands its research topics to social analysis, ubiquitous computing,\nsearch & IR, privacy and society, etc.\nThere have been", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "8e1b103b-1d79-485b-8bbe-3161ba6db37f": {"__data__": {"id_": "8e1b103b-1d79-485b-8bbe-3161ba6db37f", "embedding": null, "metadata": {"page_label": "2", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the difference between a static and dynamic heterogeneous graph?\n2. How can the dynamic perspective of heterogeneous graphs be explored and studied?\n3. What are some recent lines of research on mining heterogeneous graphs?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer model, which is used to handle multiple types of relations between different types of nodes in real-world networks. The model accounts for the timestamp of edges and the possibility of multiple timestamps for a node through the dynamic heterogeneous graph model. The section also explains the difference between the first and upcoming editions of the \"WWW\" conference node in the dynamic heterogeneous graph model.", "section_summary": "The section discusses heterogeneous graphs, which are graphs that contain nodes and edges of different types. The section explains the difference between static and dynamic heterogeneous graphs and how the dynamic perspective of heterogeneous graphs can be explored and studied. The section also discusses recent lines of research on mining heterogeneous graphs, including node classification, clustering, ranking, and representation learning. The section concludes by discussing the success of graph neural networks for relational data.", "excerpt_keywords": "1. Graph Neural Networks,\n2. Heterogeneous Graphs,\n3. Node Classification,\n4. Clustering,\n5. Ranking,\n6. Representation Learning,\n7. Relational Data,\n8. Computation Graph,\n9. Dynamic Perspective,\n10. Graph Structure."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "a3e1e120-6732-4b4a-89ee-477130998e0e", "node_type": "4", "metadata": {"page_label": "2", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "0165fd3447625d17d4424f764da78f21b9aab377a01e12e0ae99a474bdcca254"}, "2": {"node_id": "ade4f093-5318-40d7-9b81-2204d26fff8e", "node_type": "1", "metadata": {"page_label": "2", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "b270217354ba21822ccd1bdd1f5f30a1c001697cdd7e48ad448a3a6a590ebb9a"}}, "hash": "7ec7bbbeae3a50b6d5b7e41c5fee182abb0d778ff4bbf8f6311b9b23e6d1fa12", "text": "at time T,Twill be assigned to\nthe edge between the paper and conference nodes. On the contrary,\ndifferent timestamps can be assigned to a node accordingly. For\nexample, the conference node \u201cWWW\u201d can be assigned any year.\nWWW @1994 means that we are considering the first edition of\nWWW, which focuses more on internet protocol and Web infras-\ntructure, while WWW @2020 means the upcoming WWW, which\nexpands its research topics to social analysis, ubiquitous computing,\nsearch & IR, privacy and society, etc.\nThere have been significant lines of research on mining heteroge-\nnous graphs, such as node classification, clustering, ranking and\nrepresentation learning [ 3,17\u201319], while the dynamic perspective\nof HGs has not been extensively explored and studied.\n2.2 Graph Neural Networks\nRecent years have witnessed the success of graph neural net-\nworks for relational data [ 7,9,22]. Generally, a GNN can be regarded\nas using the input graph structure as the computation graph for", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "3760465e-0f22-4a23-9bef-025caa471241": {"__data__": {"id_": "3760465e-0f22-4a23-9bef-025caa471241", "embedding": null, "metadata": {"page_label": "3", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the general GNN framework and what are the most important GNN operators?\n2. What are some examples of homogeneous GNN architectures and how do they differ from each other?\n3. What is the difference between GCN, GraphSAGE, and GAT, and how do they differ in their aggregation operations?", "prev_section_summary": "The section discusses heterogeneous graphs, which are graphs that contain nodes and edges of different types. The section explains the difference between static and dynamic heterogeneous graphs and how the dynamic perspective of heterogeneous graphs can be explored and studied. The section also discusses recent lines of research on mining heterogeneous graphs, including node classification, clustering, ranking, and representation learning. The section concludes by discussing the success of graph neural networks for relational data.", "section_summary": "The section discusses the general Graph Neural Network (GNN) framework and its most important operators, Extract and Aggregate. It also describes some examples of homogeneous GNN architectures, including Graph Convolutional Network (GCN), GraphSAGE, and Graph Attention Network (GAT), and their differences in aggregation operations. The section also mentions the Heterogeneous Graph Transformer, a WWW '20 conference paper that discusses message passing in GNNs.", "excerpt_keywords": "GNN, message passing, node representation, GNN framework, Extract, Aggregate, GCN, GraphSAGE, GAT, attention mechanism, neighborhood information, graph convolutional network, linear projection, non-linear activation, RNN unit, sum, max, pooling, normalization, utility, fairness, positivity, heterogeneous graph transformer, WWW '20, Taipei, Taiwan."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "1dd47884-7047-44eb-8ed3-82f98752a400", "node_type": "4", "metadata": {"page_label": "3", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "22cd93d4a5ea32e00f61c8b2fc56465b6622bbb9cc3219da94235742c5c249dc"}, "3": {"node_id": "4bf4e7f0-65c0-4f05-a1ee-fd3e3a61809b", "node_type": "1", "metadata": {"page_label": "3", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "b43cf58a04705f81a57445813a2619c4424de5b3b22223e0c361c50401b5cb08"}}, "hash": "6bc5272bd6eb89a4abff207b1237b02ccdaa68eb469d6a2fdbc1c7156af9ece8", "text": "Heterogeneous Graph Transformer WWW \u201920, April 20\u201324, 2020, Taipei, Taiwan\nmessage passing [ 6], during which the local neighborhood informa-\ntion is aggregated to get a more contextual representation. Formally,\nit has the following form:\nDefinition 2. General GNN Framework: Suppose Hl[t]is the\nnode representation of node tat the(l)-th GNN layer, the update\nprocedure from the (l-1)-th layer to the(l)-th layer is:\nHl[t]\u2190 Aggregate\n\u2200s\u2208N(t),\u2200e\u2208E(s,t)\u0012\nExtract\u0010\nHl\u22121[s];Hl\u22121[t],e\u0011\u0013\n(1)\nwhere N(t)denotes all the source nodes of node tandE(s,t)denotes\nall the edges from node stot.\nThe most important GNN operators are Extract( \u00b7) and Aggregate(\u00b7).\nExtract(\u00b7) represents the neighbor information extractor. It extract\nuseful information from source node\u2019s representation Hl\u22121[s], with\nthe target node\u2019s representation Hl\u22121[t]and the edge ebetween the\ntwo nodes as query. Aggregate( \u00b7) gather the neighborhood informa-\ntion of souce nodes via some aggregation operators, such as mean,\nsum, andmax, while more sophisticated pooling and normalization\nfunctions can be also designed.\nVarious (homogeneous) GNN architectures have been proposed\nfollowing this framework. Kipf et al. [9] propose graph convolu-\ntional network (GCN), which averages the one-hop neighbor of each\nnode in the graph, followed by a linear projection and non-linear\nactivation operations. Hamilton et al. propose GraphSAGE that\ngeneralizes GCN\u2019s aggregation operation from average tosum, max\nand a RNN unit . Velickovi et al. propose graph attention network\n(GAT) [ 22] by introducing the attention mechanism into GNNs,\nwhich allows GAT to assign different importance to nodes within\nthe same", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "4bf4e7f0-65c0-4f05-a1ee-fd3e3a61809b": {"__data__": {"id_": "4bf4e7f0-65c0-4f05-a1ee-fd3e3a61809b", "embedding": null, "metadata": {"page_label": "3", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the difference between graph convolutional network (GCN), GraphSAGE, and graph attention network (GAT)?\n2. How do heterogeneous graph neural networks (HGNNs) differ from vanilla GNNs and GAT models?\n3. What is the purpose of parameter sharing in heterogeneous graph neural networks (HGNNs)?", "prev_section_summary": "The section discusses the general Graph Neural Network (GNN) framework and its most important operators, Extract and Aggregate. It also describes some examples of homogeneous GNN architectures, including Graph Convolutional Network (GCN), GraphSAGE, and Graph Attention Network (GAT), and their differences in aggregation operations. The section also mentions the Heterogeneous Graph Transformer, a WWW '20 conference paper that discusses message passing in GNNs.", "section_summary": "The section discusses the differences between graph convolutional network (GCN), GraphSAGE, and graph attention network (GAT) models for graph neural networks (GNNs). It also introduces heterogeneous graph neural networks (HGNNs) and their differences from vanilla GNNs and GAT models. The section highlights the limitations of previous methods for modeling heterogeneous graphs and proposes parameter sharing as a solution for better generalization. The key topics and entities discussed in the section include GCN, GraphSAGE, GAT, HGNNs, vanilla GNNs, GAT models, parameter sharing, and heterogeneous graphs.", "excerpt_keywords": "graph neural networks, GNNs, heterogeneous graphs, knowledge graphs, attention mechanism, linear projection, non-linear activation, multi-modal features, meta-path-defined edges, parameter sharing, generalization."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "1dd47884-7047-44eb-8ed3-82f98752a400", "node_type": "4", "metadata": {"page_label": "3", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "22cd93d4a5ea32e00f61c8b2fc56465b6622bbb9cc3219da94235742c5c249dc"}, "2": {"node_id": "3760465e-0f22-4a23-9bef-025caa471241", "node_type": "1", "metadata": {"page_label": "3", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "6bc5272bd6eb89a4abff207b1237b02ccdaa68eb469d6a2fdbc1c7156af9ece8"}, "3": {"node_id": "1354b19a-c9b9-44fe-825b-246ecf80ea48", "node_type": "1", "metadata": {"page_label": "3", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "18e7f737c5dcb40a9e703da8ff1f6e2f2bd1bd9825aacc31e7d39b312027ab76"}}, "hash": "b43cf58a04705f81a57445813a2619c4424de5b3b22223e0c361c50401b5cb08", "text": "this framework. Kipf et al. [9] propose graph convolu-\ntional network (GCN), which averages the one-hop neighbor of each\nnode in the graph, followed by a linear projection and non-linear\nactivation operations. Hamilton et al. propose GraphSAGE that\ngeneralizes GCN\u2019s aggregation operation from average tosum, max\nand a RNN unit . Velickovi et al. propose graph attention network\n(GAT) [ 22] by introducing the attention mechanism into GNNs,\nwhich allows GAT to assign different importance to nodes within\nthe same neighborhood.\n2.3 Heterogeneous GNNs\nRecently, studies have attempted to extend GNNs for modeling\nheterogeneous graphs. Schlichtkrull et al. [14] propose the rela-\ntional graph convolutional networks (RGCN) to model knowledge\ngraphs. RGCN keeps a distinct linear projection weight for each\nedge type. Zhang et al. [27] present the heterogeneous graph neural\nnetworks (HetGNN) that adopts different RNNs for different node\ntypes to integrate multi-modal features. Wang et al. [23] extend\ngraph attention networks by maintaining different weights for dif-\nferent meta-path-defined edges. They also use high-level semantic\nattention to differentiate and aggregate information from different\nmeta paths.\nThough these methods have shown to be empirically better than\nthe vanilla GCN and GAT models, they have not fully utilized the\nheterogeneous graphs\u2019 properties. All of them use either node type\nor edge type alone to determine GNN weight matrices. However,\nthe node or edge counts of different types can vary greatly. For\nrelations that don\u2019t have sufficient occurrences, it\u2019s hard to learn\naccurate relation-specific weights. To address this, we propose to\nconsider parameter sharing for a better generalization. Given an\nedge e=(s,t)with its meta relation as \u27e8\u03c4(s),\u03d5(e),\u03c4(t)\u27e9, if we use\nthree interaction matrices to model the three corresponding", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "1354b19a-c9b9-44fe-825b-246ecf80ea48": {"__data__": {"id_": "1354b19a-c9b9-44fe-825b-246ecf80ea48", "embedding": null, "metadata": {"page_label": "3", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the main problem that the proposed Heterogeneous Graph Transformer (HGT) addresses in the field of graph neural networks (GNNs)?\n2. How does the HGT model use parameter sharing and meta relations to improve the generalization of GNN weight matrices for heterogeneous graphs?\n3. What are the key differences between HGT and existing attempts in the field of GNNs, and how does HGT address the limitations of these existing approaches?", "prev_section_summary": "The section discusses the differences between graph convolutional network (GCN), GraphSAGE, and graph attention network (GAT) models for graph neural networks (GNNs). It also introduces heterogeneous graph neural networks (HGNNs) and their differences from vanilla GNNs and GAT models. The section highlights the limitations of previous methods for modeling heterogeneous graphs and proposes parameter sharing as a solution for better generalization. The key topics and entities discussed in the section include GCN, GraphSAGE, GAT, HGNNs, vanilla GNNs, GAT models, parameter sharing, and heterogeneous graphs.", "section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) model, which addresses the problem of limited occurrences of certain relations in heterogeneous graphs. The model uses parameter sharing and meta relations to improve the generalization of GNN weight matrices. The key differences between HGT and existing attempts in the field of GNNs include the use of meta relations to decompose interaction matrices, the incorporation of high-order heterogeneous neighbor information, the consideration of the dynamic nature of heterogeneous graphs, and the experimentation with Web-scale graphs.", "excerpt_keywords": "1. Heterogeneous Graph Transformer, 2. Transformer-like attention, 3. Meta relation, 4. Parameter sharing, 5. Attention mechanism, 6. Heterogeneous neighbor information, 7. Relative temporal encoding, 8. Web-scale graphs, 9. Graph neural networks, 10. Neural architecture."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "1dd47884-7047-44eb-8ed3-82f98752a400", "node_type": "4", "metadata": {"page_label": "3", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "22cd93d4a5ea32e00f61c8b2fc56465b6622bbb9cc3219da94235742c5c249dc"}, "2": {"node_id": "4bf4e7f0-65c0-4f05-a1ee-fd3e3a61809b", "node_type": "1", "metadata": {"page_label": "3", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "b43cf58a04705f81a57445813a2619c4424de5b3b22223e0c361c50401b5cb08"}, "3": {"node_id": "00a98235-a28f-4d0d-8435-9396a1e73830", "node_type": "1", "metadata": {"page_label": "3", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "9c0f619fee7c510fbac431d2aa90f7e3975d4faf6ee604af3f65e8df529ac559"}}, "hash": "18e7f737c5dcb40a9e703da8ff1f6e2f2bd1bd9825aacc31e7d39b312027ab76", "text": "All of them use either node type\nor edge type alone to determine GNN weight matrices. However,\nthe node or edge counts of different types can vary greatly. For\nrelations that don\u2019t have sufficient occurrences, it\u2019s hard to learn\naccurate relation-specific weights. To address this, we propose to\nconsider parameter sharing for a better generalization. Given an\nedge e=(s,t)with its meta relation as \u27e8\u03c4(s),\u03d5(e),\u03c4(t)\u27e9, if we use\nthree interaction matrices to model the three corresponding ele-\nments\u03c4(s),\u03d5(e), and\u03c4(t)in the meta relation, then the majority of\nweights could be shared. For example, in \u201cthe first author of\u201d and\n\u201cthe second author of\u201d relationships, their source and target node\ntypes are both author topaper , respectively. In other words, theknowledge about author andpaper learned from one relation could\nbe quickly transferred and adapted to the other one. Therefore, we\nintegrate this idea with the powerful Transformer-like attention\narchitecture, and propose Heterogeneous Graph Transformer.\nTo summarize, the key differences between HGT and existing\nattempts include:\n(1)Instead of attending on node or edge type alone, we use the\nmeta relation\u27e8\u03c4(s),\u03d5(e),\u03c4(t)\u27e9to decompose the interaction\nand transform matrices, enabling HGT to capture both the\ncommon and specific patterns of different relationships using\nequal or even fewer parameters.\n(2)Different from most of the existing works that are based on\ncustomized meta paths, we rely on the nature of the neural\narchitecture to incorporate high-order heterogeneous neigh-\nbor information, which automatically learns the importance\nof implicit meta paths.\n(3)Most previous works don\u2019t take the dynamic nature of (het-\nerogeneous) graphs into consideration, while we propose\nthe relative temporal encoding technique to incorporate tem-\nporal information by using limited computational resources.\n(4)None of the existing heterogeneous GNNs are designed for\nand experimented with Web-scale graphs,", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "00a98235-a28f-4d0d-8435-9396a1e73830": {"__data__": {"id_": "00a98235-a28f-4d0d-8435-9396a1e73830", "embedding": null, "metadata": {"page_label": "3", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Graph Transformer (HGT) and how does it differ from existing heterogeneous graph neural networks (GNNs)?\n2. How does the relative temporal encoding mechanism in HGT incorporate temporal information into the model?\n3. What is the overall architecture of HGT and how does it extract information from source nodes to get a contextualized representation for target nodes?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) model, which addresses the problem of limited occurrences of certain relations in heterogeneous graphs. The model uses parameter sharing and meta relations to improve the generalization of GNN weight matrices. The key differences between HGT and existing attempts in the field of GNNs include the use of meta relations to decompose interaction matrices, the incorporation of high-order heterogeneous neighbor information, the consideration of the dynamic nature of heterogeneous graphs, and the experimentation with Web-scale graphs.", "section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT), a neural network architecture for processing heterogeneous graphs. The HGT differs from existing heterogeneous graph neural networks (GNNs) by using meta relations to parameterize weight matrices for attention, message passing, and propagation steps. The section also introduces a relative temporal encoding mechanism to incorporate network dynamics and a heterogeneous mini-batch graph sampling algorithm for training on Web-scale graphs. The overall architecture of HGT consists of three components: Heterogeneous Mutual Attention, Heterogeneous Message Passing, and Target-Specific Aggregation.", "excerpt_keywords": "1. Heterogeneous Graph Transformer\n2. Meta relations\n3. Weight matrices\n4. Heterogeneous mutual attention\n5. Message passing\n6. Propagation steps\n7. Relative temporal encoding\n8. Network dynamics\n9. Sampled heterogeneous sub-graph\n10. Node representations"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "1dd47884-7047-44eb-8ed3-82f98752a400", "node_type": "4", "metadata": {"page_label": "3", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "22cd93d4a5ea32e00f61c8b2fc56465b6622bbb9cc3219da94235742c5c249dc"}, "2": {"node_id": "1354b19a-c9b9-44fe-825b-246ecf80ea48", "node_type": "1", "metadata": {"page_label": "3", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "18e7f737c5dcb40a9e703da8ff1f6e2f2bd1bd9825aacc31e7d39b312027ab76"}, "3": {"node_id": "b5b9b8c1-06ce-4c13-8251-cbc1faf88485", "node_type": "1", "metadata": {"page_label": "3", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "e87f3e99d6ae50113ae9da4af76bf2d898a4070ca04f5fd00d4b13b3a5ca7edc"}}, "hash": "9c0f619fee7c510fbac431d2aa90f7e3975d4faf6ee604af3f65e8df529ac559", "text": "from most of the existing works that are based on\ncustomized meta paths, we rely on the nature of the neural\narchitecture to incorporate high-order heterogeneous neigh-\nbor information, which automatically learns the importance\nof implicit meta paths.\n(3)Most previous works don\u2019t take the dynamic nature of (het-\nerogeneous) graphs into consideration, while we propose\nthe relative temporal encoding technique to incorporate tem-\nporal information by using limited computational resources.\n(4)None of the existing heterogeneous GNNs are designed for\nand experimented with Web-scale graphs, we therefore pro-\npose the heterogeneous Mini-Batch graph sampling algo-\nrithm designed for Web-scale graph training, enabling ex-\nperiments on the billion-scale Open Academic Graph.\n3 HETEROGENEOUS GRAPH TRANSFORMER\nIn this section, we present the Heterogeneous Graph Trans-\nformer (HGT). Its idea is to use the meta relations of heteroge-\nneous graphs to parameterize weight matrices for the heteroge-\nneous mutual attention, message passing, and propagation steps.\nTo further incorporate network dynamics, we introduce a relative\ntemporal encoding mechanism into the model.\n3.1 Overall HGT Architecture\nFigure 2 shows the overall architecture of Heterogeneous Graph\nTransformer. Given a sampled heterogeneous sub-graph (Cf. Sec-\ntion 4), HGT extracts all linked node pairs, where target node tis\nlinked by source node svia edge e. The goal of HGT is to aggregate\ninformation from source nodes to get a contextualized representa-\ntion for target node t. Such process can be decomposed into three\ncomponents: Heterogeneous Mutual Attention ,Heterogeneous Mes-\nsage Passing andTarget-Specific Aggregation .\nWe denote the output of the (l)-th HGT layer as H(l), which is\nalso the input of the (l+1)-th layer. By stacking Llayers, we can get\nthe node representations of the whole graph H(L), which can be\nused for end-to-end training or fed into downstream tasks.\n3.2 Heterogeneous Mutual Attention\nThe", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "b5b9b8c1-06ce-4c13-8251-cbc1faf88485": {"__data__": {"id_": "b5b9b8c1-06ce-4c13-8251-cbc1faf88485", "embedding": null, "metadata": {"page_label": "3", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Graph Transformer model?\n2. How does the Heterogeneous Mutual Attention component work in the HGT model?\n3. What is the difference between Heterogeneous Mutual Attention and Heterogeneous Message Passing in the HGT model?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT), a neural network architecture for processing heterogeneous graphs. The HGT differs from existing heterogeneous graph neural networks (GNNs) by using meta relations to parameterize weight matrices for attention, message passing, and propagation steps. The section also introduces a relative temporal encoding mechanism to incorporate network dynamics and a heterogeneous mini-batch graph sampling algorithm for training on Web-scale graphs. The overall architecture of HGT consists of three components: Heterogeneous Mutual Attention, Heterogeneous Message Passing, and Target-Specific Aggregation.", "section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) model, which is a multi-modal graph neural network used for corporate relative valuation. The model consists of three components: Heterogeneous Mutual Attention, Heterogeneous Message Passing, and Target-Specific Aggregation. The HGT model calculates the mutual attention between source and target nodes, and uses this information to aggregate messages from the source nodes to the target node. The model can be stacked to obtain node representations of the whole graph, which can be used for end-to-end training or fed into downstream tasks. The section also provides a brief introduction to attention-based graph neural networks.", "excerpt_keywords": "1. Heterogeneous Graph Neural Networks (HGT)\n2. Node representations\n3. Mutual attention\n4. Heterogeneous message passing\n5. Target-specific aggregation\n6. Graph convolution\n7. End-to-end training\n8. Downstream tasks\n9. Positive content\n10. Ethical considerations"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "1dd47884-7047-44eb-8ed3-82f98752a400", "node_type": "4", "metadata": {"page_label": "3", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "22cd93d4a5ea32e00f61c8b2fc56465b6622bbb9cc3219da94235742c5c249dc"}, "2": {"node_id": "00a98235-a28f-4d0d-8435-9396a1e73830", "node_type": "1", "metadata": {"page_label": "3", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "9c0f619fee7c510fbac431d2aa90f7e3975d4faf6ee604af3f65e8df529ac559"}}, "hash": "e87f3e99d6ae50113ae9da4af76bf2d898a4070ca04f5fd00d4b13b3a5ca7edc", "text": "representa-\ntion for target node t. Such process can be decomposed into three\ncomponents: Heterogeneous Mutual Attention ,Heterogeneous Mes-\nsage Passing andTarget-Specific Aggregation .\nWe denote the output of the (l)-th HGT layer as H(l), which is\nalso the input of the (l+1)-th layer. By stacking Llayers, we can get\nthe node representations of the whole graph H(L), which can be\nused for end-to-end training or fed into downstream tasks.\n3.2 Heterogeneous Mutual Attention\nThe first step is to calculate the mutual attention between source\nnode sand target node t. We first give a brief introduction to the\ngeneral attention-based GNNs as follows:\nHl[t]\u2190 Aggregate\n\u2200s\u2208N(t),\u2200e\u2208E(s,t)\u0010\nAttention(s,t)\u00b7Message(s)\u0011\n(2)", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "96e72751-1642-4479-9d54-b5a803ecee88": {"__data__": {"id_": "96e72751-1642-4479-9d54-b5a803ecee88", "embedding": null, "metadata": {"page_label": "4", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the overall architecture of the Heterogeneous Graph Transformer model and how does it work?\n2. How does the Heterogeneous Graph Transformer model handle heterogeneous graphs with different node types and their corresponding meta relations?\n3. What are the three basic operators used in the Heterogeneous Graph Transformer model and how do they work?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) model, which is a multi-modal graph neural network used for corporate relative valuation. The model consists of three components: Heterogeneous Mutual Attention, Heterogeneous Message Passing, and Target-Specific Aggregation. The HGT model calculates the mutual attention between source and target nodes, and uses this information to aggregate messages from the source nodes to the target node. The model can be stacked to obtain node representations of the whole graph, which can be used for end-to-end training or fed into downstream tasks. The section also provides a brief introduction to attention-based graph neural networks.", "section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) model, which is a multi-modal graph neural network used for corporate relative valuation. The model works by taking a sampled heterogeneous sub-graph with a target node, source nodes, and their corresponding meta relations as input and learning contextualized representations for each node. The HGT model includes three components: meta relation-aware heterogeneous mutual attention, heterogeneous message passing from source nodes, and target-specific heterogeneous message aggregation. The three basic operators used in the HGT model are Attention, Message, and Aggregate. The Graph Attention Network (GAT) is mentioned as an example of a model that uses the Attention operator, but it is noted that GAT assumes that all nodes have the same feature distributions, which is usually incorrect for heterogeneous graphs.", "excerpt_keywords": "1. Heterogeneous Graph Transformer, 2. Graph Attention Network, 3. Meta relation-aware heterogeneous mutual attention, 4. Heterogeneous message passing, 5. Target-specific heterogeneous message aggregation, 6. Attention, 7. Message, 8. Aggregate, 9. Node type, 10. HGT"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "56f63095-98b7-46bf-a7dc-790d001587c3", "node_type": "4", "metadata": {"page_label": "4", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "8277539a9bb2bc291202182cd75f19bcdff308635b52308fc69e837d85013297"}, "3": {"node_id": "2d8ae984-b0ca-4c47-b83c-b9c253d2a67f", "node_type": "1", "metadata": {"page_label": "4", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "79ba961989126a75f4be747b2578a6fd4d1e8cc78377f4ef3aae0aa1a22335d8"}}, "hash": "ce4575eb24ac235736a67539b15792cb4228361d4b736d06607e5a3ddd3b6b36", "text": "WWW \u201920, April 20\u201324, 2020, Taipei, Taiwan Ziniu Hu, Yuxiao Dong, Kuansan Wang, and Yizhou Sun\nFigure 2: The Overall Architecture of Heterogeneous Graph Transformer. Given a sampled heterogeneous sub-graph with tas\nthe target node, s1&s2as source nodes, the HGT model takes its edges e1=(s1,t)&e2=(s2,t)and their corresponding meta relations\n<\u03c4(s1),\u03d5(e1),\u03c4(t)>&<\u03c4(s2),\u03d5(e2),\u03c4(t)>as input to learn a contextualized representation H(L)for each node, which can be used for\ndownstream tasks. Color decodes the node type. HGT includes three components: (1) meta relation-aware heterogeneous mutual attention,\n(2) heterogeneous message passing from source nodes, and (3) target-specific heterogeneous message aggregation.\nwhere there are three basic operators: Attention , which estimates\nthe importance of each source node; Message , which extracts the\nmessage by using only the source node s; and Aggregate , which\naggregates the neighborhood message by the attention weight.\nFor example, the Graph Attention Network (GAT) [ 22] adopts\nan additive mechanism as Attention , uses the same weight for\ncalculating Message , and leverages the simple average followed by\na nonlinear activation for the Aggregate step. Formally, GAT has\nAttention GAT(s,t)=Softmax\n\u2200s\u2208N(t)\u0012\n\u00aea\u0010\nW Hl\u22121[t]\u2225W Hl\u22121[s]\u0011\u0013\nMessageGAT(s)=W Hl\u22121[s]\nAggregateGAT(\u00b7)=\u03c3\u0010\nMean(\u00b7)\u0011\nThough GAT is effective to give high attention values to important\nnodes, it assumes that sandthave the same feature distributions by\nusing one weight matrix W. Such an assumption, as we\u2019ve discussed\nin Section 1, is usually incorrect for heterogeneous graphs, where\neach type of nodes can have its own", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "2d8ae984-b0ca-4c47-b83c-b9c253d2a67f": {"__data__": {"id_": "2d8ae984-b0ca-4c47-b83c-b9c253d2a67f", "embedding": null, "metadata": {"page_label": "4", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation?\n2. How does the Heterogeneous Multi-Modal Graph Neural Network address the limitation of the Graph Attention mechanism for heterogeneous graphs?\n3. What is the difference between the vanilla Transformer and the proposed architecture for the Heterogeneous Multi-Modal Graph Neural Network?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) model, which is a multi-modal graph neural network used for corporate relative valuation. The model works by taking a sampled heterogeneous sub-graph with a target node, source nodes, and their corresponding meta relations as input and learning contextualized representations for each node. The HGT model includes three components: meta relation-aware heterogeneous mutual attention, heterogeneous message passing from source nodes, and target-specific heterogeneous message aggregation. The three basic operators used in the HGT model are Attention, Message, and Aggregate. The Graph Attention Network (GAT) is mentioned as an example of a model that uses the Attention operator, but it is noted that GAT assumes that all nodes have the same feature distributions, which is usually incorrect for heterogeneous graphs.", "section_summary": "The section discusses the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation. The purpose of this network is to address the limitation of the Graph Attention mechanism for heterogeneous graphs. The section explains how the Heterogeneous Multi-Modal Graph Neural Network calculates mutual attention grounded by meta relations, and how it uses distinct sets of projection weights for each meta relation. The section also compares the vanilla Transformer with the proposed architecture for the Heterogeneous Multi-Modal Graph Neural Network.", "excerpt_keywords": "1. Heterogeneous Graphs\n2. Mutual Attention\n3. Transformer\n4. Feature Distributions\n5. Meta Relations\n6. Query Vector\n7. Key Vector\n8. Dot Product\n9. Interaction Operators\n10. Parameter Sharing"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "56f63095-98b7-46bf-a7dc-790d001587c3", "node_type": "4", "metadata": {"page_label": "4", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "8277539a9bb2bc291202182cd75f19bcdff308635b52308fc69e837d85013297"}, "2": {"node_id": "96e72751-1642-4479-9d54-b5a803ecee88", "node_type": "1", "metadata": {"page_label": "4", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "ce4575eb24ac235736a67539b15792cb4228361d4b736d06607e5a3ddd3b6b36"}, "3": {"node_id": "c1eb1480-19ca-4c60-aced-2a66c976cc36", "node_type": "1", "metadata": {"page_label": "4", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "19665b49b584e3fbb912c6a6ea21a6c93e8ac8675f6160930ffde8c131a9bf04"}}, "hash": "79ba961989126a75f4be747b2578a6fd4d1e8cc78377f4ef3aae0aa1a22335d8", "text": "Hl\u22121[t]\u2225W Hl\u22121[s]\u0011\u0013\nMessageGAT(s)=W Hl\u22121[s]\nAggregateGAT(\u00b7)=\u03c3\u0010\nMean(\u00b7)\u0011\nThough GAT is effective to give high attention values to important\nnodes, it assumes that sandthave the same feature distributions by\nusing one weight matrix W. Such an assumption, as we\u2019ve discussed\nin Section 1, is usually incorrect for heterogeneous graphs, where\neach type of nodes can have its own feature distribution.\nIn view of this limitation, we design the Heterogeneous Mu-\ntual Attention mechanism. Given a target node t, and all its neigh-\nbors s\u2208N(t), which might belong to different distributions, we\nwant to calculate their mutual attention grounded by their meta\nrelations , i.e., the\u27e8\u03c4(s),\u03d5(e),\u03c4(t)\u27e9triplets.\nInspired by the architecture design of Transformer [ 21], we map\ntarget node tinto a Query vector, and source node sinto a Key vec-\ntor, and calculate their dot product as attention. The key difference\nis that the vanilla Transformer uses a single set of projections for all\nwords, while in our case each meta relation should have a distinct\nset of projection weights. To maximize parameter sharing while\nstill maintaining the specific characteristics of different relations,we propose to parameterize the weight matrices of the interac-\ntion operators into a source node projection, an edge projection,\nand a target node projection. Specifically, we calculate the h-head\nattention for each edge e=(s,t)(See Figure 2 (1)) by:\nAttention", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "c1eb1480-19ca-4c60-aced-2a66c976cc36": {"__data__": {"id_": "c1eb1480-19ca-4c60-aced-2a66c976cc36", "embedding": null, "metadata": {"page_label": "4", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the key difference between the vanilla Transformer and the proposed Heterogeneous Graph Transformer?\n2. How does the Heterogeneous Graph Transformer parameterize the weight matrices of the interaction operators to maintain specific characteristics of different relations?\n3. What is the formula for calculating the h-head attention for each edge e=(s,t) in the Heterogeneous Graph Transformer?", "prev_section_summary": "The section discusses the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation. The purpose of this network is to address the limitation of the Graph Attention mechanism for heterogeneous graphs. The section explains how the Heterogeneous Multi-Modal Graph Neural Network calculates mutual attention grounded by meta relations, and how it uses distinct sets of projection weights for each meta relation. The section also compares the vanilla Transformer with the proposed architecture for the Heterogeneous Multi-Modal Graph Neural Network.", "section_summary": "The section discusses the Heterogeneous Graph Transformer, a multi-modal graph neural network for corporate relative valuation. The key difference between the vanilla Transformer and the proposed Heterogeneous Graph Transformer is that the vanilla Transformer uses a single set of projections for all words, while the Heterogeneous Graph Transformer parameterizes the weight matrices of the interaction operators into a source node projection, an edge projection, and a target node projection. The formula for calculating the h-head attention for each edge e=(s,t) in the Heterogeneous Graph Transformer is provided, along with an explanation of how the attention is calculated by projecting the source and target nodes into the attention heads and calculating the dot product of the attention heads as attention. The section also mentions the importance of maximizing parameter sharing while still maintaining the specific characteristics of different relations.", "excerpt_keywords": "Transformer, attention, meta relations, parameter sharing, source node projection, edge projection, target node projection, h-head attention, dot product, utility, fairness, positivity, harmful, unethical, prejudiced, negative content."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "56f63095-98b7-46bf-a7dc-790d001587c3", "node_type": "4", "metadata": {"page_label": "4", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "8277539a9bb2bc291202182cd75f19bcdff308635b52308fc69e837d85013297"}, "2": {"node_id": "2d8ae984-b0ca-4c47-b83c-b9c253d2a67f", "node_type": "1", "metadata": {"page_label": "4", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "79ba961989126a75f4be747b2578a6fd4d1e8cc78377f4ef3aae0aa1a22335d8"}, "3": {"node_id": "93fc5a11-ba77-4314-a9aa-69af38413333", "node_type": "1", "metadata": {"page_label": "4", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "40b1ab85b3965bad6b7a4efe924b66c3a5b2ae92f33092c98ef1d7efe0ba84d3"}}, "hash": "19665b49b584e3fbb912c6a6ea21a6c93e8ac8675f6160930ffde8c131a9bf04", "text": "and calculate their dot product as attention. The key difference\nis that the vanilla Transformer uses a single set of projections for all\nwords, while in our case each meta relation should have a distinct\nset of projection weights. To maximize parameter sharing while\nstill maintaining the specific characteristics of different relations,we propose to parameterize the weight matrices of the interac-\ntion operators into a source node projection, an edge projection,\nand a target node projection. Specifically, we calculate the h-head\nattention for each edge e=(s,t)(See Figure 2 (1)) by:\nAttention HGT(s,e,t)=Softmax\n\u2200s\u2208N(t)\u0010\n\u2225\ni\u2208[1,h]ATT-headi(s,e,t)\u0011\n(3)\nATT-headi(s,e,t)=\u0010\nKi(s)WATT\n\u03d5(e)Qi(t)T\u0011\n\u00b7\u00b5\u27e8\u03c4(s),\u03d5(e),\u03c4(t)\u27e9\u221a\nd\nKi(s)=K-Lineari\n\u03c4(s)\u0010\nH(l\u22121)[s]\u0011\nQi(t)=Q-Lineari\n\u03c4(t)\u0010\nH(l\u22121)[t]\u0011\nFirst, for the i-th attention head ATT-headi(s,e,t), we project the\n\u03c4(s)-type source node sinto the i-thKeyvector Ki(s)with a linear\nprojection K-Lineari\n\u03c4(s):Rd\u2192Rd\nh, where his the number of\nattention heads andd\nhis the vector dimension per head. Note that\nK-Lineari\n\u03c4(s)is indexed by the source node s\u2019s type\u03c4(s), meaning\nthat each type of nodes has a unique linear projection to maximally\nmodel the distribution differences. Similarly, we also project the\ntarget node twith a linear projection Q-Lineari\n\u03c4(t)into the i\u2212th\nQuery vector.\nNext, we need to calculate the similarity between the Query\nvector", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "93fc5a11-ba77-4314-a9aa-69af38413333": {"__data__": {"id_": "93fc5a11-ba77-4314-a9aa-69af38413333", "embedding": null, "metadata": {"page_label": "4", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using attention heads and vector dimensions in the Heterogeneous Graph Transformer model?\n2. How does the model capture different semantic relations between node types in heterogeneous graphs?\n3. What is the role of the WATT \u03d5(e) matrix in the Heterogeneous Graph Transformer model?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer, a multi-modal graph neural network for corporate relative valuation. The key difference between the vanilla Transformer and the proposed Heterogeneous Graph Transformer is that the vanilla Transformer uses a single set of projections for all words, while the Heterogeneous Graph Transformer parameterizes the weight matrices of the interaction operators into a source node projection, an edge projection, and a target node projection. The formula for calculating the h-head attention for each edge e=(s,t) in the Heterogeneous Graph Transformer is provided, along with an explanation of how the attention is calculated by projecting the source and target nodes into the attention heads and calculating the dot product of the attention heads as attention. The section also mentions the importance of maximizing parameter sharing while still maintaining the specific characteristics of different relations.", "section_summary": "The section discusses the Heterogeneous Graph Transformer model, which is a type of multi-modal graph neural network used for corporate relative valuation. The model uses attention heads and vector dimensions to capture different semantic relations between node types in heterogeneous graphs. The WATT \u03d5(e) matrix is also discussed, which is used to capture different semantic relations between node types. The model is designed to capture the unique characteristics of heterogeneous graphs, such as different edge types between node pairs.", "excerpt_keywords": "1. Transformer,\n2. Heterogeneous graphs,\n3. Linear projections,\n4. Edge types,\n5. Dot product,\n6. Edge-based matrix,\n7. Semantic relations,\n8. Node types,\n9. Query vector,\n10. Key vector."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "56f63095-98b7-46bf-a7dc-790d001587c3", "node_type": "4", "metadata": {"page_label": "4", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "8277539a9bb2bc291202182cd75f19bcdff308635b52308fc69e837d85013297"}, "2": {"node_id": "c1eb1480-19ca-4c60-aced-2a66c976cc36", "node_type": "1", "metadata": {"page_label": "4", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "19665b49b584e3fbb912c6a6ea21a6c93e8ac8675f6160930ffde8c131a9bf04"}}, "hash": "40b1ab85b3965bad6b7a4efe924b66c3a5b2ae92f33092c98ef1d7efe0ba84d3", "text": "where his the number of\nattention heads andd\nhis the vector dimension per head. Note that\nK-Lineari\n\u03c4(s)is indexed by the source node s\u2019s type\u03c4(s), meaning\nthat each type of nodes has a unique linear projection to maximally\nmodel the distribution differences. Similarly, we also project the\ntarget node twith a linear projection Q-Lineari\n\u03c4(t)into the i\u2212th\nQuery vector.\nNext, we need to calculate the similarity between the Query\nvector Qi(t)and Key vector Ki(s). One unique characteristic of\nheterogeneous graphs is that there may exist different edge types\n(relations) between a node type pair, e.g., \u03c4(s)and\u03c4(t). Therefore,\nunlike the vanilla Transformer that directly calculates the dot prod-\nuct between the Query and Key vectors, we keep a distinct edge-\nbased matrix WATT\n\u03d5(e)\u2208Rd\nh\u00d7d\nhfor each edge type \u03d5(e). In doing so,\nthe model can capture different semantic relations even between", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "c46ef647-4874-4fbc-be8f-80e3c1b4dcd5": {"__data__": {"id_": "c46ef647-4874-4fbc-be8f-80e3c1b4dcd5", "embedding": null, "metadata": {"page_label": "5", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Graph Transformer model in the context of corporate relative valuation?\n2. How does the model incorporate meta-relations of edges into the attention and message passing processes?\n3. What is the role of the prior tensor \u00b5 in the Heterogeneous Graph Transformer model?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer model, which is a type of multi-modal graph neural network used for corporate relative valuation. The model uses attention heads and vector dimensions to capture different semantic relations between node types in heterogeneous graphs. The WATT \u03d5(e) matrix is also discussed, which is used to capture different semantic relations between node types. The model is designed to capture the unique characteristics of heterogeneous graphs, such as different edge types between node pairs.", "section_summary": "The section discusses the Heterogeneous Graph Transformer model, which is used for corporate relative valuation. The model incorporates meta-relations of edges into the attention and message passing processes. The prior tensor \u00b5 is used to denote the general significance of each meta-relation triplet and serves as an adaptive scaling to the attention. The section also explains the Heterogeneous Message Passing process, which incorporates the meta-relations of edges into the message passing process to alleviate the distribution differences of nodes and edges of different types.", "excerpt_keywords": "1. Heterogeneous Graph Transformer, 2. Attention, 3. Message Passing, 4. Meta Relations, 5. Node Pairs, 6. Target Nodes, 7. Adaptive Scaling, 8. Softmax, 9. Multi-Head Message, 10. Edge Dependency"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "142f0092-e09f-411b-81fc-dd5b2fe76ffb", "node_type": "4", "metadata": {"page_label": "5", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "864598e743cc8b86a13d888c7f1fe9f382b30358a7cb9f69741e3d9ed7e533bf"}, "3": {"node_id": "eccc65c3-7dd9-4b2d-a4ef-4e495f0c90c9", "node_type": "1", "metadata": {"page_label": "5", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "3143a62e3b9937ef5d08796020c1639e440cc1b3e62d06ade284705cc62a5a14"}}, "hash": "66b47a8caccc8fb81645ab707a21c1617797d58270c5f8b0dcf79d63fcf8ffa0", "text": "Heterogeneous Graph Transformer WWW \u201920, April 20\u201324, 2020, Taipei, Taiwan\nthe same node type pairs. Moreover, since not all the relation-\nships contribute equally to the target nodes, we add a prior tensor\n\u00b5\u2208R|A|\u00d7|R|\u00d7|A|to denote the general significance of each meta\nrelation triplet, serving as an adaptive scaling to the attention.\nFinally, we concatenate hattention heads together to get the\nattention vector for each node pair. Then, for each target node t,\nwe gather all attention vectors from its neighbors N(t)and conduct\nsoftmax, making it fulfill\u00cd\n\u2200s\u2208N(t)Attention HGT(s,e,t)=1h\u00d71.\n3.3 Heterogeneous Message Passing\nParallel to the calculation of mutual attention, we pass informa-\ntion from source nodes to target nodes (See Figure 2 (2)). Similar\nto the attention process, we would like to incorporate the meta\nrelations of edges into the message passing process to alleviate the\ndistribution differences of nodes and edges of different types. For a\npair of nodes e=(s,t), we calculate its multi-head Message by:\nMessageHGT(s,e,t)=\u2225\ni\u2208[1,h]MSG -headi(s,e,t) (4)\nMSG -headi(s,e,t)=M-Lineari\n\u03c4(s)\u0010\nH(l\u22121)[s]\u0011\nWMSG\n\u03d5(e)\nTo get the i-th message head MSG -headi(s,e,t), we first project the\n\u03c4(s)-type source node sinto the i-th message vector with a linear\nprojection M-Lineari\n\u03c4(s):Rd\u2192Rd\nh. It is then followed by a matrix\nWMSG\n\u03d5(e)\u2208Rd\nh\u00d7d\nhfor incorporating the edge dependency. The final\nstep is to concat all hmessage heads to get the MessageHGT(s,e,t)\nfor each", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "eccc65c3-7dd9-4b2d-a4ef-4e495f0c90c9": {"__data__": {"id_": "eccc65c3-7dd9-4b2d-a4ef-4e495f0c90c9", "embedding": null, "metadata": {"page_label": "5", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Graph Transformer model for Corporate Relative Valuation?\n2. How does the Heterogeneous Multi-Head Attention and Message Calculation process work in the Heterogeneous Graph Transformer model?\n3. What is the Target-Specific Aggregation process in the Heterogeneous Graph Transformer model and how does it work?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer model, which is used for corporate relative valuation. The model incorporates meta-relations of edges into the attention and message passing processes. The prior tensor \u00b5 is used to denote the general significance of each meta-relation triplet and serves as an adaptive scaling to the attention. The section also explains the Heterogeneous Message Passing process, which incorporates the meta-relations of edges into the message passing process to alleviate the distribution differences of nodes and edges of different types.", "section_summary": "The section discusses the Heterogeneous Graph Transformer model for Corporate Relative Valuation. The model uses heterogeneous multi-head attention and message calculation to aggregate information from source nodes to the target node. The Target-Specific Aggregation process is used to map the updated vector back to its type-specific distribution. The section also mentions the small-world property of real-world graphs and how it affects the model's output.", "excerpt_keywords": "1. Heterogeneous graph, 2. Multi-head attention, 3. Message passing, 4. Linear projection, 5. Matrix multiplication, 6. Edge dependency, 7. Node pair, 8. Target-specific aggregation, 9. Type-specific distribution, 10. Residual connection."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "142f0092-e09f-411b-81fc-dd5b2fe76ffb", "node_type": "4", "metadata": {"page_label": "5", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "864598e743cc8b86a13d888c7f1fe9f382b30358a7cb9f69741e3d9ed7e533bf"}, "2": {"node_id": "c46ef647-4874-4fbc-be8f-80e3c1b4dcd5", "node_type": "1", "metadata": {"page_label": "5", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "66b47a8caccc8fb81645ab707a21c1617797d58270c5f8b0dcf79d63fcf8ffa0"}, "3": {"node_id": "fc8cafcd-884b-4f2c-868b-21b9b442fd04", "node_type": "1", "metadata": {"page_label": "5", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "021c9882d98a9fe1f6527e8e3a00e260d2f3aa3772cf6a00712116f810704b0c"}}, "hash": "3143a62e3b9937ef5d08796020c1639e440cc1b3e62d06ade284705cc62a5a14", "text": "get the i-th message head MSG -headi(s,e,t), we first project the\n\u03c4(s)-type source node sinto the i-th message vector with a linear\nprojection M-Lineari\n\u03c4(s):Rd\u2192Rd\nh. It is then followed by a matrix\nWMSG\n\u03d5(e)\u2208Rd\nh\u00d7d\nhfor incorporating the edge dependency. The final\nstep is to concat all hmessage heads to get the MessageHGT(s,e,t)\nfor each node pair.\n3.4 Target-Specific Aggregation\nWith the heterogeneous multi-head attention and message cal-\nculated, we need to aggregate them from the source nodes to the\ntarget node (See Figure 2 (3)). Note that the softmax procedure in\nEq. 3 has made the sum of each target node t\u2019s attention vectors\nto one, we can thus simply use the attention vector as the weight\nto average the corresponding messages from the source nodes and\nget the updated vector eH(l)[t]as:\neH(l)[t]=\u2295\n\u2200s\u2208N(t)\u0010\nAttention HGT(s,e,t)\u00b7MessageHGT(s,e,t)\u0011\n.\nThis aggregates information to the target node tfrom all its neigh-\nbors (source nodes) of different feature distributions.\nThe final step is to map target node t\u2019s vector back to its type-\nspecific distribution, indexed by its node type \u03c4(t). To do so, we\napply a linear projection A-Linear \u03c4(t)to the updated vector eH(l)[t],\nfollowed by residual connection [8] as:\nH(l)[t]=A-Linear \u03c4(t)\u0010\n\u03c3\u0000eH(l)[t]\u0001\u0011\n+H(l\u22121)[t]. (5)\nIn this way, we get the l-th HGT layer\u2019s output H(l)[t]for the target\nnode t. Due to the \u201csmall-world\u201d property of real-world graphs,\nstacking the", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "fc8cafcd-884b-4f2c-868b-21b9b442fd04": {"__data__": {"id_": "fc8cafcd-884b-4f2c-868b-21b9b442fd04", "embedding": null, "metadata": {"page_label": "5", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using a linear projection and residual connection in the Heterogeneous Graph Transformer model?\n2. How does the \"small-world\" property of real-world graphs enable the HGT model to generate highly contextualized representations for each node?\n3. What is the role of the meta relation in parameterizing the weight matrices in the HGT model, and how does it improve the model's capacity and efficiency compared to existing models?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer model for Corporate Relative Valuation. The model uses heterogeneous multi-head attention and message calculation to aggregate information from source nodes to the target node. The Target-Specific Aggregation process is used to map the updated vector back to its type-specific distribution. The section also mentions the small-world property of real-world graphs and how it affects the model's output.", "section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) model, which is a multi-modal graph neural network used for corporate relative valuation. The model uses a linear projection and residual connection to generate highly contextualized representations for each node. The \"small-world\" property of real-world graphs enables the model to generate these representations. The meta relation is used to parameterize the weight matrices in the model, which improves its capacity and efficiency compared to existing models. The section also discusses the Relative Temporal Encoding (RTE) process, which models graph dynamic.", "excerpt_keywords": "1. Heterogeneous Graphs, 2. Graph Neural Networks, 3. Hierarchical Graph Transformers, 4. Relative Temporal Encoding, 5. Node Classification, 6. Link Prediction, 7. Small-World Property, 8. Meta Relation, 9. Parameterization, 10. Efficiency"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "142f0092-e09f-411b-81fc-dd5b2fe76ffb", "node_type": "4", "metadata": {"page_label": "5", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "864598e743cc8b86a13d888c7f1fe9f382b30358a7cb9f69741e3d9ed7e533bf"}, "2": {"node_id": "eccc65c3-7dd9-4b2d-a4ef-4e495f0c90c9", "node_type": "1", "metadata": {"page_label": "5", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "3143a62e3b9937ef5d08796020c1639e440cc1b3e62d06ade284705cc62a5a14"}, "3": {"node_id": "f271ba5c-b90d-41b8-ba61-abd944580ccb", "node_type": "1", "metadata": {"page_label": "5", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "b25f2599583729d2fa72026d3c4a204515eea6a632d29948776a3a542872c099"}}, "hash": "021c9882d98a9fe1f6527e8e3a00e260d2f3aa3772cf6a00712116f810704b0c", "text": "a linear projection A-Linear \u03c4(t)to the updated vector eH(l)[t],\nfollowed by residual connection [8] as:\nH(l)[t]=A-Linear \u03c4(t)\u0010\n\u03c3\u0000eH(l)[t]\u0001\u0011\n+H(l\u22121)[t]. (5)\nIn this way, we get the l-th HGT layer\u2019s output H(l)[t]for the target\nnode t. Due to the \u201csmall-world\u201d property of real-world graphs,\nstacking the HGT blocks for Llayers ( Lbeing a small value) can\nenable each node reaching a large proportion of nodes\u2014with differ-\nent types and relations\u2014in the full graph. That is, HGT generates\na highly contextualized representation H(L)for each node, which\ncan be fed into any models to conduct downstream heterogeneous\nnetwork tasks, such as node classification and link prediction.\nFigure 3: Relative Temporal Encoding (RTE) to model graph\ndynamic. Nodes are associated with timestamps T(\u00b7). After the\nRTE process, the temporal augmented representations are fed to\nthe HGT model.\nThrough the whole model architecture, we highly rely on using\nthemeta relation \u2014\u27e8\u03c4(s),\u03d5(e),\u03c4(t)\u27e9\u2014to parameterize the weight\nmatrices separately. This can be interpreted as a trade-off between\nthe model capacity and efficiency. Compared with the vanilla Trans-\nformer, our model distinguishes the operators for different relations\nand thus is more capable to handle the distribution differences in\nheterogeneous graphs. Compared with existing models that keep a\ndistinct matrix for each meta relation as a whole, HGT\u2019s triplet pa-\nrameterization can better leverage the heterogeneous graph schema\nto achieve parameter sharing. On one hand, relations with few oc-\ncurrences can benefit from such parameter sharing for fast adapta-\ntion and generalization. On the other hand, different relationships\u2019\noperators can still maintain their specific characteristics by using a\nmuch smaller parameter set.\n3.5 Relative Temporal Encoding\nBy far, we", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "f271ba5c-b90d-41b8-ba61-abd944580ccb": {"__data__": {"id_": "f271ba5c-b90d-41b8-ba61-abd944580ccb", "embedding": null, "metadata": {"page_label": "5", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of Heterogeneous Graph Transformer (HGT) and how does it differ from existing models for modeling heterogeneous graphs?\n2. What is the Relative Temporal Encoding (RTE) technique and how does it handle dynamic information in heterogeneous graphs?\n3. How does the parameter sharing in HGT's triplet parameterization help in achieving faster adaptation and generalization for relations with few occurrences, while still maintaining the specific characteristics of different relationships' operators?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) model, which is a multi-modal graph neural network used for corporate relative valuation. The model uses a linear projection and residual connection to generate highly contextualized representations for each node. The \"small-world\" property of real-world graphs enables the model to generate these representations. The meta relation is used to parameterize the weight matrices in the model, which improves its capacity and efficiency compared to existing models. The section also discusses the Relative Temporal Encoding (RTE) process, which models graph dynamic.", "section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) model for modeling heterogeneous graphs. HGT differs from existing models by using triplet parameterization for parameter sharing, which helps in achieving faster adaptation and generalization for relations with few occurrences while still maintaining the specific characteristics of different relationships' operators. The section also introduces the Relative Temporal Encoding (RTE) technique for handling dynamic information in heterogeneous graphs, which is inspired by Transformer's positional encoding method. RTE allows nodes and edges with different timestamps to interact with each other and model the dynamic dependencies in heterogeneous graphs.", "excerpt_keywords": "graph neural network, heterogeneous graphs, parameter sharing, relative temporal encoding, graph dynamic, Transformer, positional encoding, sequential dependencies, long texts, time slots."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "142f0092-e09f-411b-81fc-dd5b2fe76ffb", "node_type": "4", "metadata": {"page_label": "5", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "864598e743cc8b86a13d888c7f1fe9f382b30358a7cb9f69741e3d9ed7e533bf"}, "2": {"node_id": "fc8cafcd-884b-4f2c-868b-21b9b442fd04", "node_type": "1", "metadata": {"page_label": "5", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "021c9882d98a9fe1f6527e8e3a00e260d2f3aa3772cf6a00712116f810704b0c"}}, "hash": "b25f2599583729d2fa72026d3c4a204515eea6a632d29948776a3a542872c099", "text": "to handle the distribution differences in\nheterogeneous graphs. Compared with existing models that keep a\ndistinct matrix for each meta relation as a whole, HGT\u2019s triplet pa-\nrameterization can better leverage the heterogeneous graph schema\nto achieve parameter sharing. On one hand, relations with few oc-\ncurrences can benefit from such parameter sharing for fast adapta-\ntion and generalization. On the other hand, different relationships\u2019\noperators can still maintain their specific characteristics by using a\nmuch smaller parameter set.\n3.5 Relative Temporal Encoding\nBy far, we present HGT\u2014a graph neural network for modeling\nheterogeneous graphs. Next, we introduce the Relative Temporal\nEncoding (RTE) technique for HGT to handle graph dynamic.\nThe traditional way to incorporate temporal information is to\nconstruct a separate graph for each time slot. However, such a pro-\ncedure may lose a large portion of structural dependencies across\ndifferent time slots. Meanwhile, the representation of a node at\ntime tmight rely on edges that happen at other time slots. There-\nfore, a proper way to model dynamic graphs is to maintain all the\nedges happening at different times and allow nodes and edges with\ndifferent timestamps to interact with each other.\nIn light of this, we propose the Relative Temporal Encoding\n(RTE) mechanism to model the dynamic dependencies in heteroge-\nneous graphs. RTE is inspired by Transformer\u2019s positional encoding\nmethod [ 15,21], which has been shown successful to capture the\nsequential dependencies of words in long texts.\nSpecifically, given a source node sand a target node t, along\nwith their corresponding timestamps T(s)andT(t), we denote the\nrelative time gap \u2206T(t,s)=T(t)\u2212T(s)as an index to get a relative\ntemporal encoding RT E(\u2206T(t,s)). Noted that the training dataset", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "6abd790d-8794-4c04-bcf0-f669a2529da4": {"__data__": {"id_": "6abd790d-8794-4c04-bcf0-f669a2529da4", "embedding": null, "metadata": {"page_label": "6", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using sinusoid functions as a basis for the temporal encoding in Heterogeneous Graph Transformer (HGT)?\n2. How does the Heterogeneous Mini-Batch Graph Sampling algorithm (HGSampling) address the scalability issue of full-batch GNN training for Web-scale heterogeneous graphs?\n3. What is the inductive timestamp assignment method used in HGT for training Web-scale heterogeneous graphs with dynamic information?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) model for modeling heterogeneous graphs. HGT differs from existing models by using triplet parameterization for parameter sharing, which helps in achieving faster adaptation and generalization for relations with few occurrences while still maintaining the specific characteristics of different relationships' operators. The section also introduces the Relative Temporal Encoding (RTE) technique for handling dynamic information in heterogeneous graphs, which is inspired by Transformer's positional encoding method. RTE allows nodes and edges with different timestamps to interact with each other and model the dynamic dependencies in heterogeneous graphs.", "section_summary": "The section discusses the use of sinusoid functions as a basis for temporal encoding in Heterogeneous Graph Transformer (HGT) and the Heterogeneous Mini-Batch Graph Sampling algorithm (HGSampling) for scalable training of Web-scale heterogeneous graphs. The inductive timestamp assignment method is also introduced for training Web-scale heterogeneous graphs with dynamic information. The section highlights the challenges of full-batch GNN training for Web-scale graphs and the need for scalable and efficient methods for training heterogeneous graphs.", "excerpt_keywords": "1. Graph Neural Networks (GNNs), 2. Heterogeneous Graphs, 3. Web-scale, 4. Dynamic Information, 5. Temporal Encoding, 6. Target Node, 7. Source Node, 8. Timestamp Assignment, 9. Mini-Batch Graph Sampling, 10. Inductive Learning"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "a6390232-e74a-4453-ac1c-f469dac6f577", "node_type": "4", "metadata": {"page_label": "6", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "c14210eb722eadddb07bb607a0a10354c9147de1afbed69ef7eebfb20d641da7"}, "3": {"node_id": "b42e822c-3234-4543-99f4-28f3def94b28", "node_type": "1", "metadata": {"page_label": "6", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "0732b78ccc98ec4631cb333d824fec8392019913e1decb95436e32aa341ab944"}}, "hash": "7c5d3613af88944cc27da24dd177df2b07b2e31ed468545c2f0ca0605a90617a", "text": "WWW \u201920, April 20\u201324, 2020, Taipei, Taiwan Ziniu Hu, Yuxiao Dong, Kuansan Wang, and Yizhou Sun\nwill not cover all possible time gaps, and thus RT E should be capable\nof generalizing to unseen times and time gaps. Therefore, we adopt\na fixed set of sinusoid functions as basis, with a tunable linear\nprojection T-Linear\u2217:Rd\u2192RdasRT E:\nBase\u0000\u2206T(t,s),2i\u0001=sin\u0010\n\u2206Tt,s/100002i\nd\u0011\n(6)\nBase\u0000\u2206T(t,s),2i+1\u0001=cos\u0010\n\u2206Tt,s/100002i+1\nd\u0011\n(7)\nRT E\u0000\u2206T(t,s)\u0001=T-Linear\u0010\nBase(\u2206Tt,s)\u0011\n(8)\nFinally, the temporal encoding relative to the target node tis added\nto the source node s\u2019 representation as follows:\nbH(l\u22121)[s]=H(l\u22121)[s]+RT E\u0000\u2206T(t,s)\u0001(9)\nIn this way, the temporal augmented representation bH(l\u22121)will\ncapture the relative temporal information of source node sand\ntarget node t. The RTE procedure is illustrated in the Figure 3.\n4 WEB-SCALE HGT TRAINING\nIn this section, we present HGT\u2019s strategies for training Web-\nscale heterogeneous graphs with dynamic information, including\nan efficient Heterogeneous Mini-Batch Graph Sampling algorithm\u2014\nHGSampling\u2014and an inductive timestamp assignment method.\n4.1 HGSampling\nThe full-batch GNN [ 9] training requires the calculation of all\nnode representations per layer, making it not scalable for Web-scale\ngraphs. To address this issue, different sampling-based methods [ 1,\n2,7,29] have been proposed to train GNNs on a subset of nodes.\nHowever, directly using them for heterogeneous graphs is prone to\nget", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "b42e822c-3234-4543-99f4-28f3def94b28": {"__data__": {"id_": "b42e822c-3234-4543-99f4-28f3def94b28", "embedding": null, "metadata": {"page_label": "6", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Mini-Batch Graph Sampling algorithm (HGSampling) and how does it address the issue of imbalanced node types in heterogeneous graphs?\n2. Can you explain the importance of the node budget and the importance sampling strategy in the HGSampling algorithm?\n3. How does the normalization of the degree in the HGSampling algorithm help to reduce variance and minimize information loss?", "prev_section_summary": "The section discusses the use of sinusoid functions as a basis for temporal encoding in Heterogeneous Graph Transformer (HGT) and the Heterogeneous Mini-Batch Graph Sampling algorithm (HGSampling) for scalable training of Web-scale heterogeneous graphs. The inductive timestamp assignment method is also introduced for training Web-scale heterogeneous graphs with dynamic information. The section highlights the challenges of full-batch GNN training for Web-scale graphs and the need for scalable and efficient methods for training heterogeneous graphs.", "section_summary": "The section discusses the Heterogeneous Mini-Batch Graph Sampling algorithm (HGSampling) for training Graph Neural Networks (GNNs) on heterogeneous graphs. The algorithm addresses the issue of imbalanced node types in heterogeneous graphs by keeping a similar number of nodes and edges for each type and minimizing information loss and sample variance. The algorithm uses a node budget and an importance sampling strategy to sample an equal number of nodes per type. The normalization of the degree in the algorithm helps to reduce variance and minimize information loss. The section also mentions the importance of efficient HGSampling for handling Web-scale heterogeneous graphs and the inductive timestamp assignment method.", "excerpt_keywords": "1. Heterogeneous Graphs\n2. Graph Neural Networks (GNNs)\n3. Sampling-based methods\n4. Web-scale graphs\n5. Node types\n6. Degree distribution\n7. Importance sampling\n8. Information loss\n9. Sample variance\n10. Heterogeneous Mini-Batch Graph Sampling (HGSampling)"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "a6390232-e74a-4453-ac1c-f469dac6f577", "node_type": "4", "metadata": {"page_label": "6", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "c14210eb722eadddb07bb607a0a10354c9147de1afbed69ef7eebfb20d641da7"}, "2": {"node_id": "6abd790d-8794-4c04-bcf0-f669a2529da4", "node_type": "1", "metadata": {"page_label": "6", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "7c5d3613af88944cc27da24dd177df2b07b2e31ed468545c2f0ca0605a90617a"}, "3": {"node_id": "6a2ab275-5181-4cc2-b554-cc30e92ebb1d", "node_type": "1", "metadata": {"page_label": "6", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "9986ca430ee342fcbbd99a1af3db3ddab2f46fe11d556c9a8a663f4e64fc3150"}}, "hash": "0732b78ccc98ec4631cb333d824fec8392019913e1decb95436e32aa341ab944", "text": "heterogeneous graphs with dynamic information, including\nan efficient Heterogeneous Mini-Batch Graph Sampling algorithm\u2014\nHGSampling\u2014and an inductive timestamp assignment method.\n4.1 HGSampling\nThe full-batch GNN [ 9] training requires the calculation of all\nnode representations per layer, making it not scalable for Web-scale\ngraphs. To address this issue, different sampling-based methods [ 1,\n2,7,29] have been proposed to train GNNs on a subset of nodes.\nHowever, directly using them for heterogeneous graphs is prone to\nget sub-graphs that are extremely imbalanced regarding different\nnode types, due to that the degree distribution and the total number\nof nodes for each type can vary dramatically.\nTo address this issue, we propose an efficient Heterogeneous\nMini-Batch Graph Sampling algorithm\u2014HGSampling\u2014to enable\nboth HGT and traditional GNNs to handle Web-scale heterogeneous\ngraphs. HGSampling is able to 1) keep a similar number of nodes\nand edges for each type and 2) keep the sampled sub-graph dense\nto minimize the information loss and reduce the sample variance.\nAlgorithm 1 outlines the HGSampling algorithm. Its basic idea\nis to keep a separate node budget B[\u03c4]for each node type \u03c4and\nto sample an equal number of nodes per type with an importance\nsampling strategy to reduce variance. Given node talready sampled,\nwe add all its direct neighbors into the corresponding budget with\nAlgorithm 2, and add t\u2019s normalized degree to these neighbors in\nline 8, which will then be used to calculate the sampling probability.\nSuch normalization is equivalent to accumulate the random walk\nprobability of each sampled node to its neighborhood, avoiding the\nsampling being dominated by high-degree nodes. Intuitively, the\nhigher such value is, the more a candidate node is correlated with\nthe currently sampled nodes, and thus should be given a higher\nprobability to be sampled.\n\u2217For simplicity, we denote a linear projection L :Ra\u2192Rbas a function to conduct\nlinear transformation to vector x\u2208Raas: L(x)=W x+b,", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "6a2ab275-5181-4cc2-b554-cc30e92ebb1d": {"__data__": {"id_": "6a2ab275-5181-4cc2-b554-cc30e92ebb1d", "embedding": null, "metadata": {"page_label": "6", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Mini-Batch Graph Sampling algorithm?\n2. How does the normalization process in the algorithm ensure fairness in the sampling of nodes?\n3. What is the role of the Budget data structure in the Heterogeneous Mini-Batch Graph Sampling algorithm?", "prev_section_summary": "The section discusses the Heterogeneous Mini-Batch Graph Sampling algorithm (HGSampling) for training Graph Neural Networks (GNNs) on heterogeneous graphs. The algorithm addresses the issue of imbalanced node types in heterogeneous graphs by keeping a similar number of nodes and edges for each type and minimizing information loss and sample variance. The algorithm uses a node budget and an importance sampling strategy to sample an equal number of nodes per type. The normalization of the degree in the algorithm helps to reduce variance and minimize information loss. The section also mentions the importance of efficient HGSampling for handling Web-scale heterogeneous graphs and the inductive timestamp assignment method.", "section_summary": "The section discusses the Heterogeneous Mini-Batch Graph Sampling algorithm, which is used to sample nodes from a heterogeneous graph. The algorithm ensures fairness in the sampling of nodes by normalizing the degree of each node and accumulating the random walk probability of each sampled node to its neighborhood. The Budget data structure is used to store nodes for each node type with normalized degree, and the algorithm samples nodes based on their probability of being sampled. The section also explains the role of the linear projection function L in the algorithm and how it conducts linear transformation to vector x.", "excerpt_keywords": "1. Graph sampling\n2. Heterogeneous mini-batch graph sampling\n3. Adjacency matrix\n4. Node set\n5. Sample number\n6. Sample depth\n7. Normalized degree\n8. Budget\n9. Linear projection\n10. Matrix W\n11. Bias b\n12. Node type\n13. Source node\n14. Sampling probability\n15. Sampled node set\n16. Sampled adjacency matrix\n17. Intuitive sampling\n18. Correlation\n19. Fairness\n20. Positivity"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "a6390232-e74a-4453-ac1c-f469dac6f577", "node_type": "4", "metadata": {"page_label": "6", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "c14210eb722eadddb07bb607a0a10354c9147de1afbed69ef7eebfb20d641da7"}, "2": {"node_id": "b42e822c-3234-4543-99f4-28f3def94b28", "node_type": "1", "metadata": {"page_label": "6", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "0732b78ccc98ec4631cb333d824fec8392019913e1decb95436e32aa341ab944"}, "3": {"node_id": "f4110556-56cf-4a3b-854c-63133116eb05", "node_type": "1", "metadata": {"page_label": "6", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "7e6dabbab3b0cc695a111ef6a9e56c5636a72027d37036d5fd662dceff805e3f"}}, "hash": "9986ca430ee342fcbbd99a1af3db3ddab2f46fe11d556c9a8a663f4e64fc3150", "text": "which will then be used to calculate the sampling probability.\nSuch normalization is equivalent to accumulate the random walk\nprobability of each sampled node to its neighborhood, avoiding the\nsampling being dominated by high-degree nodes. Intuitively, the\nhigher such value is, the more a candidate node is correlated with\nthe currently sampled nodes, and thus should be given a higher\nprobability to be sampled.\n\u2217For simplicity, we denote a linear projection L :Ra\u2192Rbas a function to conduct\nlinear transformation to vector x\u2208Raas: L(x)=W x+b, where matrix W\u2208Ra+b\nand bias b\u2208Rb.Wandbare learnable parameters for L.Algorithm 1 Heterogeneous Mini-Batch Graph Sampling\nRequire: Adjacency matrix Afor each\u27e8\u03c4(s),\u03d5(e),\u03c4(t)\u27e9relation\npair; Output node Set OS; Sample number nper node type;\nSample depth L.\nEnsure: Sampled node set NS; Sampled adjacency matrix \u02c6A.\n1:NS\u2190OS// Initialize sampled node set as output node set.\n2:Initialize an empty Budget Bstoring nodes for each node type\nwith normalized degree.\n3:fort\u2208NSdo\n4: Add-In-Budget( B,t,A,NS) // Add neighbors of ttoB.\n5:end for\n6:forl\u21901toLdo\n7:forsource node type \u03c4\u2208Bdo\n8: forsource node s\u2208B[\u03c4]do\n9: prob(l\u22121)[\u03c4][s]\u2190B[\u03c4][s]2\n\u2225B[\u03c4]\u22252\n2// Calculate sampling prob-\nability for each source node sof node type \u03c4.\n10: end for\n11: Sample nnodes{ti}n\ni=1from B[\u03c4]using prob(l\u22121)[\u03c4].\n12: fort\u2208{ti}n\ni=1do\n13: OS[\u03c4].add(t)// Add node tinto Output node", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "f4110556-56cf-4a3b-854c-63133116eb05": {"__data__": {"id_": "f4110556-56cf-4a3b-854c-63133116eb05", "embedding": null, "metadata": {"page_label": "6", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Graph Transformer algorithm and how does it reduce sampling variance?\n2. How does the algorithm ensure that the sampled sub-graph is sufficiently dense and contains a similar number of nodes per type?\n3. What is the role of the normalized degree and importance sampling in reducing sampling variance in the Heterogeneous Graph Transformer algorithm?", "prev_section_summary": "The section discusses the Heterogeneous Mini-Batch Graph Sampling algorithm, which is used to sample nodes from a heterogeneous graph. The algorithm ensures fairness in the sampling of nodes by normalizing the degree of each node and accumulating the random walk probability of each sampled node to its neighborhood. The Budget data structure is used to store nodes for each node type with normalized degree, and the algorithm samples nodes based on their probability of being sampled. The section also explains the role of the linear projection function L in the algorithm and how it conducts linear transformation to vector x.", "section_summary": "The section discusses the Heterogeneous Graph Transformer algorithm, which is used to reduce sampling variance in training Graph Neural Networks (GNNs) on Web-scale heterogeneous graphs. The algorithm involves calculating sampling probabilities for each node based on their normalized degree and importance sampling, and then sampling nodes from the budget using these probabilities. The resulting sampled sub-graph is sufficiently dense and contains a similar number of nodes per type, making it suitable for training GNNs. The section also mentions the importance of timestamp assignment in GNN training, but does not discuss it in detail.", "excerpt_keywords": "1. Sampling, 2. Graph, 3. Neural Networks, 4. Heterogeneous, 5. Adjacency Matrix, 6. Budget, 7. Importance Sampling, 8. Normalized Degree, 9. Variance, 10. Web-scale"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "a6390232-e74a-4453-ac1c-f469dac6f577", "node_type": "4", "metadata": {"page_label": "6", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "c14210eb722eadddb07bb607a0a10354c9147de1afbed69ef7eebfb20d641da7"}, "2": {"node_id": "6a2ab275-5181-4cc2-b554-cc30e92ebb1d", "node_type": "1", "metadata": {"page_label": "6", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "9986ca430ee342fcbbd99a1af3db3ddab2f46fe11d556c9a8a663f4e64fc3150"}, "3": {"node_id": "09cb40e9-ec9e-400f-9eef-1708cefdcd04", "node_type": "1", "metadata": {"page_label": "6", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "850ddb33e0e729f6e3ac7204c485c2dbe27fc947cce59ea4fda17eb99be4b6e2"}}, "hash": "7e6dabbab3b0cc695a111ef6a9e56c5636a72027d37036d5fd662dceff805e3f", "text": "node s\u2208B[\u03c4]do\n9: prob(l\u22121)[\u03c4][s]\u2190B[\u03c4][s]2\n\u2225B[\u03c4]\u22252\n2// Calculate sampling prob-\nability for each source node sof node type \u03c4.\n10: end for\n11: Sample nnodes{ti}n\ni=1from B[\u03c4]using prob(l\u22121)[\u03c4].\n12: fort\u2208{ti}n\ni=1do\n13: OS[\u03c4].add(t)// Add node tinto Output node set.\n14: Add-In-Budget( B,t,A,NS) // Add neighbors of ttoB.\n15: B[\u03c4].pop(t)// Remove sampled node tfrom Budget.\n16: end for\n17: end for\n18:end for\n19:Reconstruct the sampled adjacency matrix \u02c6Aamong the sam-\npled nodes OSfrom A.\n20:return OSand \u02c6A;\nAfter the budget is updated, we then calculate the sampling\nprobability in Algorithm 1 line 9, where we calculate the square of\nthe cumulative normalized degree of each node sin each budget.\nAs proved in [ 29], using such sampling probability can reduce the\nsampling variance. Then, we sample nnodes in type \u03c4by using the\ncalculated probability, add them into the output node set, update\nits neighborhood to the budget, and remove it out of the budget\nin lines 12\u201315. Repeating such procedure for Ltimes, we get a\nsampled sub-graph with Ldepth from the initial nodes. Finally, we\nreconstruct the adjacency matrix among the sampled nodes. By\nusing the above algorithm, the sampled sub-graph contains a similar\nnumber of nodes per type (based on the separate node budget), and\nis sufficiently dense to reduce the sampling variance (based on the\nnormalized degree and importance sampling), making it suitable\nfor training GNNs on Web-scale heterogeneous graphs.\n4.2 Inductive Timestamp Assignment\nTill now we have assumed that each node tis assigned with\na timestamp T(t). However,", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "09cb40e9-ec9e-400f-9eef-1708cefdcd04": {"__data__": {"id_": "09cb40e9-ec9e-400f-9eef-1708cefdcd04", "embedding": null, "metadata": {"page_label": "6", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the algorithm used to sample a sub-graph from the initial nodes in the Heterogeneous Graph Transformer model?\n2. How does the algorithm ensure that the sampled sub-graph is suitable for training Graph Neural Networks on Web-scale heterogeneous graphs?\n3. What is the purpose of assigning timestamps to nodes in real-world heterogeneous graphs, and how is it done in the Heterogeneous Graph Transformer model?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer algorithm, which is used to reduce sampling variance in training Graph Neural Networks (GNNs) on Web-scale heterogeneous graphs. The algorithm involves calculating sampling probabilities for each node based on their normalized degree and importance sampling, and then sampling nodes from the budget using these probabilities. The resulting sampled sub-graph is sufficiently dense and contains a similar number of nodes per type, making it suitable for training GNNs. The section also mentions the importance of timestamp assignment in GNN training, but does not discuss it in detail.", "section_summary": "The section discusses the Heterogeneous Graph Transformer model, which is used for training Graph Neural Networks (GNNs) on Web-scale heterogeneous graphs. The algorithm used to sample a sub-graph from the initial nodes in the model is described, as well as the purpose of assigning timestamps to nodes in real-world heterogeneous graphs. The section also explains the concept of plain nodes, which are nodes that are not associated with a fixed time, and the need for inductive timestamp assignment to decide which timestamp(s) to attach to these nodes.", "excerpt_keywords": "1. Heterogeneous graphs, 2. Sampling, 3. Adjacency matrix, 4. Graph neural networks, 5. Web-scale, 6. Timestamp assignment, 7. Plain nodes, 8. Inductive, 9. Importance sampling, 10. Normalized degree"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "a6390232-e74a-4453-ac1c-f469dac6f577", "node_type": "4", "metadata": {"page_label": "6", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "c14210eb722eadddb07bb607a0a10354c9147de1afbed69ef7eebfb20d641da7"}, "2": {"node_id": "f4110556-56cf-4a3b-854c-63133116eb05", "node_type": "1", "metadata": {"page_label": "6", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "7e6dabbab3b0cc695a111ef6a9e56c5636a72027d37036d5fd662dceff805e3f"}}, "hash": "850ddb33e0e729f6e3ac7204c485c2dbe27fc947cce59ea4fda17eb99be4b6e2", "text": "sub-graph with Ldepth from the initial nodes. Finally, we\nreconstruct the adjacency matrix among the sampled nodes. By\nusing the above algorithm, the sampled sub-graph contains a similar\nnumber of nodes per type (based on the separate node budget), and\nis sufficiently dense to reduce the sampling variance (based on the\nnormalized degree and importance sampling), making it suitable\nfor training GNNs on Web-scale heterogeneous graphs.\n4.2 Inductive Timestamp Assignment\nTill now we have assumed that each node tis assigned with\na timestamp T(t). However, in real-world heterogeneous graphs,\nmany nodes are not associated with a fixed time. Therefore, we\nneed to assign different timestamps to it. We denote these nodes as\nplain nodes. For example, the WWW conference is held in both 1974\nand 2019, and the WWW node in these two years has dramatically\ndifferent research topics. Consequently, we need to decide which\ntimestamp(s) to attach to the WWW node.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "004441c6-45e6-4480-93d6-4f85f41c09d3": {"__data__": {"id_": "004441c6-45e6-4480-93d6-4f85f41c09d3", "embedding": null, "metadata": {"page_label": "7", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Graph Transformer algorithm and how does it work?\n2. How does the Add-In-Budget algorithm work in the context of Heterogeneous Graph Transformer?\n3. What is the role of timestamp assignment in the Heterogeneous Graph Transformer algorithm and how does it work?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer model, which is used for training Graph Neural Networks (GNNs) on Web-scale heterogeneous graphs. The algorithm used to sample a sub-graph from the initial nodes in the model is described, as well as the purpose of assigning timestamps to nodes in real-world heterogeneous graphs. The section also explains the concept of plain nodes, which are nodes that are not associated with a fixed time, and the need for inductive timestamp assignment to decide which timestamp(s) to attach to these nodes.", "section_summary": "The section discusses the Heterogeneous Graph Transformer algorithm, which is used for corporate relative valuation. The algorithm works by sampling nodes from a heterogeneous graph and assigning them timestamps based on event nodes they are linked with. The Add-In-Budget algorithm is also introduced, which updates the budget for each node type based on the normalized degree of the added node t. The section also mentions the role of timestamp assignment in the Heterogeneous Graph Transformer algorithm and how it works.", "excerpt_keywords": "1. Heterogeneous Graphs\n2. Timestamp Assignment\n3. Inductive Timestamping\n4. Event Nodes\n5. Publication Behavior\n6. Publication Date\n7. Plan Nodes\n8. Temporal Dependency\n9. Normalized Degree\n10. Budget Allocation"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "bb57f321-c60f-4e25-8444-a5cdbee87b27", "node_type": "4", "metadata": {"page_label": "7", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "c2df2dcb54ce0499daf86ff9549dbd4f0bbdb90573278a8d17f7b8b45eb9301f"}, "3": {"node_id": "7ea36911-fd07-4f09-b748-1cb9ca04f495", "node_type": "1", "metadata": {"page_label": "7", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "4fe2ca6d0ece35dde4a95d42f529f4f2f4e0691834d57b6bd2360621d8a9692d"}}, "hash": "8c9a03184f7b5c4e41a5a600a28042793359ce7f4f980577b47e4f98caa3f054", "text": "Heterogeneous Graph Transformer WWW \u201920, April 20\u201324, 2020, Taipei, Taiwan\nFigure 4: HGSampling with Inductive Timestamp Assignment.\nAlgorithm 2 Add-In-Budget\nRequire: Budget Bstoring nodes for each type with normal-\nized degree; Added node t; Adjacency matrix Afor each\n\u27e8\u03c4(s),\u03d5(e),\u03c4(t)\u27e9relation pair; Sampled node set NS.\nEnsure: Updated Budget B.\n1:foreach possible source node type \u03c4and edge type \u03d5do\n2: \u02c6Dt\u21901/len\u0010\nA\u27e8\u03c4,\u03d5,\u03c4(t)\u27e9[t]\u0011\n// get normalized degree of\nadded node tregarding to\u27e8\u03c4,\u03d5,\u03c4(t)\u27e9.\n3:forsource node sinA\u27e8\u03c4,\u03d5,\u03c4(t)\u27e9[t]do\n4: ifshas not been sampled ( s<NS)then\n5: ifshas no timestamp then\n6: s.time =t.time // Inductively inherit timestamp.\n7: end if\n8: B[\u03c4][s]\u2190 B[\u03c4][s]+\u02c6Dt// Add candidate node sto\nbudget Bwith target node t\u2019s normalized degree.\n9: end if\n10: end for\n11:end for\n12:return Updated Budget B\nThere also exist event nodes in heterogeneous graphs that have\nan explicit timestamp associated with them. For example, the pa-\nper node should be associated with its publication behavior and\ntherefore attached to its publication date.\nWe propose an inductive timestamp assignment algorithm to\nassign plain nodes timestamps based on event nodes that they are\nlinked with. The algorithm is shown in Algorithm 2 line 6. The idea\nis that plan nodes inherit the timestamps from event nodes. We\nexamine whether the candidate source node is an event node. If\nyes, like a paper published at a specific year, we keep its timestamp\nfor capturing temporal dependency. If no, like a conference", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "7ea36911-fd07-4f09-b748-1cb9ca04f495": {"__data__": {"id_": "7ea36911-fd07-4f09-b748-1cb9ca04f495", "embedding": null, "metadata": {"page_label": "7", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the timestamp assignment algorithm proposed in the paper \"Heterogeneous Graph Transformer\"?\n2. How does the proposed timestamp assignment algorithm adaptively assign timestamps during the sub-graph sampling procedure?\n3. What are the three heterogeneous academic graph datasets used for evaluation in the paper \"Heterogeneous Graph Transformer\"?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer algorithm, which is used for corporate relative valuation. The algorithm works by sampling nodes from a heterogeneous graph and assigning them timestamps based on event nodes they are linked with. The Add-In-Budget algorithm is also introduced, which updates the budget for each node type based on the normalized degree of the added node t. The section also mentions the role of timestamp assignment in the Heterogeneous Graph Transformer algorithm and how it works.", "section_summary": "The section discusses the purpose and implementation of a timestamp assignment algorithm proposed in the paper \"Heterogeneous Graph Transformer\". The algorithm assigns timestamps to plain nodes based on event nodes they are linked with, allowing for adaptive timestamping during the sub-graph sampling procedure. The section also describes the evaluation of the proposed model on three heterogeneous academic graph datasets, including the Open Academic Graph, and the Paper-Field, Paper-Venue, and Author Disambiguation tasks. Additionally, the section mentions the ability of the model to automatically learn and extract meta paths important for downstream tasks.", "excerpt_keywords": "1. Heterogeneous Graph Transformer\n2. Inductive timestamp assignment algorithm\n3. Paper-Field prediction\n4. Paper-Venue prediction\n5. Author Disambiguation\n6. Open Academic Graph (OAG)\n7. Web-scale datasets\n8. Domain-specific subgraphs\n9. Generalization\n10. Citation links"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "bb57f321-c60f-4e25-8444-a5cdbee87b27", "node_type": "4", "metadata": {"page_label": "7", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "c2df2dcb54ce0499daf86ff9549dbd4f0bbdb90573278a8d17f7b8b45eb9301f"}, "2": {"node_id": "004441c6-45e6-4480-93d6-4f85f41c09d3", "node_type": "1", "metadata": {"page_label": "7", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "8c9a03184f7b5c4e41a5a600a28042793359ce7f4f980577b47e4f98caa3f054"}, "3": {"node_id": "20458bfb-7830-4b2e-a62f-8e15a800029c", "node_type": "1", "metadata": {"page_label": "7", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "312033805cdf725f4b2885acb3ff6e2afac83fa5abb5117332f09a4752456ca4"}}, "hash": "4fe2ca6d0ece35dde4a95d42f529f4f2f4e0691834d57b6bd2360621d8a9692d", "text": "with them. For example, the pa-\nper node should be associated with its publication behavior and\ntherefore attached to its publication date.\nWe propose an inductive timestamp assignment algorithm to\nassign plain nodes timestamps based on event nodes that they are\nlinked with. The algorithm is shown in Algorithm 2 line 6. The idea\nis that plan nodes inherit the timestamps from event nodes. We\nexamine whether the candidate source node is an event node. If\nyes, like a paper published at a specific year, we keep its timestamp\nfor capturing temporal dependency. If no, like a conference that\ncan be associated with any timestamp, we inductively assign the\nassociated node\u2019s timestamp, such as the published year of its paper,to this plain node. In this way, we can adaptively assign timestamps\nduring the sub-graph sampling procedure.\n5 EVALUATION\nIn this section, we evaluate the proposed Heterogeneous Graph\nTransformer on three heterogeneous academic graph datasets. We\nconduct the Paper-Field prediction, Paper-Venue prediction, and\nAuthor Disambiguation tasks. We also take case studies to demon-\nstrate how HGT can automatically learn and extract meta paths\nthat are important for downstream tasks\u2020.\n5.1 Web-Scale Datasets\nTo examine the performance of the proposed model and its real-\nworld applications, we use the Open Academic Graph (OAG) [ 16,\n20,28] as our experimental basis. OAG consists of more than 178\nmillion nodes and 2.236 billion edges\u2014the largest publicly available\nheterogeneous academic dataset. In addition, all papers in OAG are\nassociated with their publication dates, spanning from 1900 to 2019.\nTo test the generalization of the proposed model, we also con-\nstruct two domain-specific subgraphs from OAG: the Computer\nScience (CS) and Medicine (Med) academic graphs. The graph sta-\ntistics are listed in Table 1, in which P\u2013A, P\u2013F, P\u2013V, A\u2013I, and P\u2013P\ndenote the edges between paper and author, paper and field, paper\nand venue, author and institute, and the citation links between", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "20458bfb-7830-4b2e-a62f-8e15a800029c": {"__data__": {"id_": "20458bfb-7830-4b2e-a62f-8e15a800029c", "embedding": null, "metadata": {"page_label": "7", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of constructing domain-specific subgraphs from OAG for testing the generalization of the proposed model?\n2. How does the size and distinguishability of the CS and Med graphs differ from other commonly used academic datasets in existing heterogeneous GNN and graph mining studies?\n3. What is the availability and location of the dataset and code for the proposed heterogeneous graph neural network model?", "prev_section_summary": "The section discusses the purpose and implementation of a timestamp assignment algorithm proposed in the paper \"Heterogeneous Graph Transformer\". The algorithm assigns timestamps to plain nodes based on event nodes they are linked with, allowing for adaptive timestamping during the sub-graph sampling procedure. The section also describes the evaluation of the proposed model on three heterogeneous academic graph datasets, including the Open Academic Graph, and the Paper-Field, Paper-Venue, and Author Disambiguation tasks. Additionally, the section mentions the ability of the model to automatically learn and extract meta paths important for downstream tasks.", "section_summary": "The section discusses the purpose of constructing domain-specific subgraphs from OAG for testing the generalization of a proposed heterogeneous graph neural network model. The size and distinguishability of the Computer Science (CS) and Medicine (Med) academic graphs are compared to other commonly used academic datasets in existing heterogeneous GNN and graph mining studies. The dataset and code for the proposed heterogeneous graph neural network model are publicly available at https://github.com/acbull/pyHGT.", "excerpt_keywords": "1. Graph Neural Networks, 2. Heterogeneous Graphs, 3. Academic Graphs, 4. Computer Science, 5. Medicine, 6. Publication Dates, 7. Node and Edge Statistics, 8. Citation Links, 9. Domain-Specific Subgraphs, 10. Large-Scale Datasets"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "bb57f321-c60f-4e25-8444-a5cdbee87b27", "node_type": "4", "metadata": {"page_label": "7", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "c2df2dcb54ce0499daf86ff9549dbd4f0bbdb90573278a8d17f7b8b45eb9301f"}, "2": {"node_id": "7ea36911-fd07-4f09-b748-1cb9ca04f495", "node_type": "1", "metadata": {"page_label": "7", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "4fe2ca6d0ece35dde4a95d42f529f4f2f4e0691834d57b6bd2360621d8a9692d"}}, "hash": "312033805cdf725f4b2885acb3ff6e2afac83fa5abb5117332f09a4752456ca4", "text": "addition, all papers in OAG are\nassociated with their publication dates, spanning from 1900 to 2019.\nTo test the generalization of the proposed model, we also con-\nstruct two domain-specific subgraphs from OAG: the Computer\nScience (CS) and Medicine (Med) academic graphs. The graph sta-\ntistics are listed in Table 1, in which P\u2013A, P\u2013F, P\u2013V, A\u2013I, and P\u2013P\ndenote the edges between paper and author, paper and field, paper\nand venue, author and institute, and the citation links between two\npapers.\nBoth the CS and Med graphs contain tens of millions of nodes\nand hundreds of millions of edges, making them at least one mag-\nnitude larger than the other CS (e.g., DBLP) and Med (e.g., Pubmed)\nacademic datasets that are commonly used in existing heteroge-\nneous GNN and heterogeneous graph mining studies. Moreover,\nthe three datasets used are far more distinguishable than previously\nwide-adopted small citation graphs used in GNN studies, such as\n\u2020The dataset and code are publicly available at https://github.com/acbull/pyHGT.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "13bf3d72-249e-4505-8136-7246fad2c783": {"__data__": {"id_": "13bf3d72-249e-4505-8136-7246fad2c783", "embedding": null, "metadata": {"page_label": "8", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation?\n2. How does the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation differentiate between the different node types and edge types in the Open Academic Graph (OAG)?\n3. What are the performance metrics used to evaluate the effectiveness of the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation on the Open Academic Graph (OAG)?", "prev_section_summary": "The section discusses the purpose of constructing domain-specific subgraphs from OAG for testing the generalization of a proposed heterogeneous graph neural network model. The size and distinguishability of the Computer Science (CS) and Medicine (Med) academic graphs are compared to other commonly used academic datasets in existing heterogeneous GNN and graph mining studies. The dataset and code for the proposed heterogeneous graph neural network model are publicly available at https://github.com/acbull/pyHGT.", "section_summary": "The section discusses the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation, which is a graph neural network that is used to predict the relative valuation of a company based on various factors such as its papers, authors, fields, venues, and institutions. The network differentiates between the different node types and edge types in the Open Academic Graph (OAG), which is a dataset that contains information about academic papers, authors, fields, venues, and institutions. The performance metrics used to evaluate the effectiveness of the network on the OAG include precision, recall, F1 score, and accuracy.", "excerpt_keywords": "1. Open Academic Graph (OAG)\n2. Node types: Paper, Author, Field, Venue, Institute\n3. Field levels: L0 to L5\n4. Author orders: first, last, others\n5. Venue types: journal, conference, preprint\n6. Self-loop connection\n7. GNN architectures\n8. Keyword extraction\n9. Information retrieval\n10. Academic network analysis"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "70efd5d5-bcda-4698-a7e4-c0ba91bf514a", "node_type": "4", "metadata": {"page_label": "8", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "70254454ab988b1a5a2393bcda5e976341fb56875518f7a867588d109b8f7365"}, "3": {"node_id": "f7527283-444c-483e-857d-e02388c28f30", "node_type": "1", "metadata": {"page_label": "8", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "4faf3793a591690022d16bf82dcac57b9bc76ff4c9edf32d6de430a321b0de02"}}, "hash": "bf380eb278353fb25605ac29f65d35beaf7c9428ff01bd33ad52af6048b63161", "text": "WWW \u201920, April 20\u201324, 2020, Taipei, Taiwan Ziniu Hu, Yuxiao Dong, Kuansan Wang, and Yizhou Sun\nDataset #nodes #edges #papers #authors #fields #venues #institutes #P-A #P-F #P-V #A-I #P-P\nCS 11,732,027 107,263,811 5,597,605 5,985,759 119,537 27,433 16,931 15,571,614 47,462,559 5,597,606 7,190,480 31,441,552\nMed 51,044,324 451,468,375 21,931,587 28,779,507 289,930 25,044 18,256 85,620,479 149,728,483 21,931,588 28,779,507 165,408,318\nOAG 178,663,927 2,236,196,802 89,606,257 88,364,081 615,228 53,073 25,288 300,853,688 657,049,405 89,606,258 167,449,933 1,021,237,518\nTable 1: Open Academic Graph (OAG) Statistics.\nCora, Citeseer, and Pubmed [ 9,22], which only contain thousands\nof nodes.\nThere are totally five node types: \u2018Paper\u2019, \u2018Author\u2019, \u2018Field\u2019, \u2018Venue\u2019,\nand \u2018Institute\u2019. The \u2018Field\u2019 nodes in OAG are categorized into six\nlevels from L0toL5, which are organized with a hierarchical tree.\nTherefore, we differentiate the \u2018Paper\u2013Field\u2019 edges corresponding\nto the field level.\nIn addition, we differentiate the different author orders (i.e., the\nfirst author, the last one, and others) and venue types (i.e., journal,\nconference, and preprint) as well. Finally, the \u2018Self\u2019 type corre-\nsponds to the self-loop connection, which is widely added in GNN\narchitectures.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "f7527283-444c-483e-857d-e02388c28f30": {"__data__": {"id_": "f7527283-444c-483e-857d-e02388c28f30", "embedding": null, "metadata": {"page_label": "8", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Graph Transformer model and what tasks does it evaluate on?\n2. How does the model differentiate between different types of edges and nodes in the OAG?\n3. What are the evaluation metrics used to measure the performance of the HGT model and how are they calculated?", "prev_section_summary": "The section discusses the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation, which is a graph neural network that is used to predict the relative valuation of a company based on various factors such as its papers, authors, fields, venues, and institutions. The network differentiates between the different node types and edge types in the Open Academic Graph (OAG), which is a dataset that contains information about academic papers, authors, fields, venues, and institutions. The performance metrics used to evaluate the effectiveness of the network on the OAG include precision, recall, F1 score, and accuracy.", "section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) model, which is a multi-modal graph neural network used for corporate relative valuation. The model differentiates between different types of edges and nodes in the OAG, including the 'Self' relationship, which is symmetric, and all other relation types have a reverse relation type. The model is evaluated on four real-world downstream tasks: Paper\u2013Field (L1), Paper\u2013Field (L2), Paper\u2013Venue, and Author Disambiguation. The evaluation metrics used are NDCG and MRR, which are widely adopted ranking metrics. The model is compared with two classes of state-of-the-art graph neural networks.", "excerpt_keywords": "1. Heterogeneous Graph Neural Networks (HGT)\n2. Node classification\n3. Paper-Field prediction\n4. Paper-Venue prediction\n5. Author disambiguation\n6. Graph neural networks (GNNs)\n7. Softmax output layer\n8. Neural Tensor Networks (NTNs)\n9. Real-world downstream tasks\n10. Evaluation metrics (NDCG, MRR)"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "70efd5d5-bcda-4698-a7e4-c0ba91bf514a", "node_type": "4", "metadata": {"page_label": "8", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "70254454ab988b1a5a2393bcda5e976341fb56875518f7a867588d109b8f7365"}, "2": {"node_id": "13bf3d72-249e-4505-8136-7246fad2c783", "node_type": "1", "metadata": {"page_label": "8", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "bf380eb278353fb25605ac29f65d35beaf7c9428ff01bd33ad52af6048b63161"}, "3": {"node_id": "4b37021e-9968-4f85-b8b9-3cb8e730ad15", "node_type": "1", "metadata": {"page_label": "8", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "05aa0d513f045c09501bc1c6acc1ada1f5cccde7c0bb58eb9b669d4ba4b4ea7b"}}, "hash": "4faf3793a591690022d16bf82dcac57b9bc76ff4c9edf32d6de430a321b0de02", "text": "nodes in OAG are categorized into six\nlevels from L0toL5, which are organized with a hierarchical tree.\nTherefore, we differentiate the \u2018Paper\u2013Field\u2019 edges corresponding\nto the field level.\nIn addition, we differentiate the different author orders (i.e., the\nfirst author, the last one, and others) and venue types (i.e., journal,\nconference, and preprint) as well. Finally, the \u2018Self\u2019 type corre-\nsponds to the self-loop connection, which is widely added in GNN\narchitectures. Except the \u2018Self\u2019 relationship, which are symmetric,\nall other relation types \u03d5have a reverse relation type \u03d5\u22121.\n5.2 Experimental Setup\nTasks and Evaluation. We evaluate the HGT model on four dif-\nferent real-world downstream tasks: the prediction of Paper\u2013Field\n(L1), Paper\u2013Field ( L2), and Paper\u2013Venue, and Author Disambigua-\ntion. The goal of the first three node classification tasks is to predict\nthe correct L1andL2fields that each paper belongs to or the venue\nit is published at, respectively. We use different GNNs to get the\ncontextual node representation of the paper and use a softmax out-\nput layer to get its classification label. For author disambiguation,\nwe select all the authors with the same name and their associated\npapers. The task is to conduct link prediction between these papers\nand candidate authors. After getting the paper and author node\nrepresentations from GNNs, we use a Neural Tensor Network to\nget the probability of each author-paper pair to be linked.\nFor all tasks, we use papers published before the year 2015 as\nthe training set, papers between 2015 and 2016 for validation, and\npapers between 2016 and 2019 as testing. We choose NDCG and\nMRR, which are two widely adopted ranking metrics [ 10,11], as\nthe evaluation metrics. All models are trained for 5 times and, the\nmean and standard variance of test performance are reported.\nBaselines. We compare HGT with two classes of state-of-art graph\nneural", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "4b37021e-9968-4f85-b8b9-3cb8e730ad15": {"__data__": {"id_": "4b37021e-9968-4f85-b8b9-3cb8e730ad15", "embedding": null, "metadata": {"page_label": "8", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using a Neural Tensor Network in the Heterogeneous Graph Transformer model?\n2. How are the evaluation metrics NDCG and MRR used to measure the performance of the HGT model?\n3. What are the differences between the two classes of GNN baselines used in the HGT model, and how do they compare to the dedicated heterogeneous GNN baselines?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) model, which is a multi-modal graph neural network used for corporate relative valuation. The model differentiates between different types of edges and nodes in the OAG, including the 'Self' relationship, which is symmetric, and all other relation types have a reverse relation type. The model is evaluated on four real-world downstream tasks: Paper\u2013Field (L1), Paper\u2013Field (L2), Paper\u2013Venue, and Author Disambiguation. The evaluation metrics used are NDCG and MRR, which are widely adopted ranking metrics. The model is compared with two classes of state-of-the-art graph neural networks.", "section_summary": "The section discusses the use of a Neural Tensor Network in the Heterogeneous Graph Transformer model to predict the probability of author-paper pairs being linked. The evaluation metrics NDCG and MRR are used to measure the performance of the model. The section compares the HGT model with two classes of state-of-the-art graph neural networks, including Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT), as well as several dedicated heterogeneous GNNs as baselines, including Relational Graph Convolutional Networks (RGCN), Heterogeneous Graph Neural Networks (HetGNN), and Heterogeneous Graph Attention Networks (HAN). The section also discusses the effectiveness of the two major components of HGT, namely heterogeneous weight parameterization.", "excerpt_keywords": "1. Heterogeneous Graph Neural Networks (HGT)\n2. Graph Convolutional Networks (GCN)\n3. Graph Attention Networks (GAT)\n4. Relational Graph Convolutional Networks (RGCN)\n5. Heterogeneous Graph Neural Networks (HetGNN)\n6. Heterogeneous Graph Attention Networks (HAN)\n7. Neural Tensor Network\n8. Probability of author-paper pair link\n9. Fairness and positivity\n10. NDCG and MRR evaluation metrics"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "70efd5d5-bcda-4698-a7e4-c0ba91bf514a", "node_type": "4", "metadata": {"page_label": "8", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "70254454ab988b1a5a2393bcda5e976341fb56875518f7a867588d109b8f7365"}, "2": {"node_id": "f7527283-444c-483e-857d-e02388c28f30", "node_type": "1", "metadata": {"page_label": "8", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "4faf3793a591690022d16bf82dcac57b9bc76ff4c9edf32d6de430a321b0de02"}, "3": {"node_id": "36bb2f32-dc7b-4109-94a8-ec50e301c6a7", "node_type": "1", "metadata": {"page_label": "8", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "9bcaafea70f31abe07fed514eb4cca9e371a25e05ad6126cf9fe9a0d1cace3ca"}}, "hash": "05aa0d513f045c09501bc1c6acc1ada1f5cccde7c0bb58eb9b669d4ba4b4ea7b", "text": "use a Neural Tensor Network to\nget the probability of each author-paper pair to be linked.\nFor all tasks, we use papers published before the year 2015 as\nthe training set, papers between 2015 and 2016 for validation, and\npapers between 2016 and 2019 as testing. We choose NDCG and\nMRR, which are two widely adopted ranking metrics [ 10,11], as\nthe evaluation metrics. All models are trained for 5 times and, the\nmean and standard variance of test performance are reported.\nBaselines. We compare HGT with two classes of state-of-art graph\nneural networks. All baselines as well as our own model, are imple-\nmented via the PyTorch Geometric (PyG) package [5].\nThe first class of GNN baselines is designed for homogeneous\ngraphs, including:\n\u2022Graph Convolutional Networks (GCN) [ 9], which simply\naverages the neighbor\u2019s embedding followed by linear pro-\njection. We use the implementation provided in PyG.\n\u2022Graph Attention Networks (GAT) [ 22], which adopts multi-\nhead additive attention on neighbors. We use the implemen-\ntation provided in PyG.\nThe second class considered is several dedicated heterogeneous\nGNNs as baselines, including:\u2022Relational Graph Convolutional Networks (RGCN) [ 14], which\nkeeps a different weight for each relationship, i.e., a relation\ntriplet. We use the implementation provided in PyG.\n\u2022Heterogeneous Graph Neural Networks (HetGNN) [ 27], which\nadopts different Bi-LSTMs for different node type for aggre-\ngating neighbor information. We re-implement this model\nin PyG following the authors\u2019 official code.\n\u2022Heterogeneous Graph Attention Networks (HAN) [ 23] de-\nsign hierarchical attentions to aggregate neighbor informa-\ntion via different meta paths. We re-implement this model\nin PyG following the authors\u2019 official code.\nIn addition, to systematically analyze the effectiveness of the\ntwo major components of HGT, i.e., Heterogeneous weight pa-\nrameterization", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "36bb2f32-dc7b-4109-94a8-ec50e301c6a7": {"__data__": {"id_": "36bb2f32-dc7b-4109-94a8-ec50e301c6a7", "embedding": null, "metadata": {"page_label": "8", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Graph Transformer model and how does it differ from traditional Graph Neural Networks?\n2. How does the HGSampling algorithm proposed in Section 4 handle large-scale OAG graphs and avoid data leakage?\n3. What are the input features used to represent each type of nodes in the Heterogeneous Graph Transformer model and how are they obtained?", "prev_section_summary": "The section discusses the use of a Neural Tensor Network in the Heterogeneous Graph Transformer model to predict the probability of author-paper pairs being linked. The evaluation metrics NDCG and MRR are used to measure the performance of the model. The section compares the HGT model with two classes of state-of-the-art graph neural networks, including Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT), as well as several dedicated heterogeneous GNNs as baselines, including Relational Graph Convolutional Networks (RGCN), Heterogeneous Graph Neural Networks (HetGNN), and Heterogeneous Graph Attention Networks (HAN). The section also discusses the effectiveness of the two major components of HGT, namely heterogeneous weight parameterization.", "section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) model, which is a type of graph neural network (GNN) designed to handle heterogeneous graphs with multiple types of nodes. The model uses a combination of heterogeneous weight parameterization and relative temporal encoding to aggregate neighbor information and make predictions. The section also describes the HGSampling algorithm proposed to handle large-scale graphs and avoid data leakage. Input features for each type of node are discussed, including using pre-trained word embeddings for paper nodes and metapath2vec for field, venue, and institute nodes. The section concludes with an ablation study comparing the effectiveness of different components of the HGT model.", "excerpt_keywords": "1. Graph attention networks\n2. Heterogeneous graph attention networks\n3. Heterogeneous weight parameterization\n4. Relative temporal encoding\n5. Heterogeneous graph attention networks\n6. Ablation study\n7. Large-scale OAG graph\n8. Data leakage\n9. Input features\n10. Node embeddings"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "70efd5d5-bcda-4698-a7e4-c0ba91bf514a", "node_type": "4", "metadata": {"page_label": "8", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "70254454ab988b1a5a2393bcda5e976341fb56875518f7a867588d109b8f7365"}, "2": {"node_id": "4b37021e-9968-4f85-b8b9-3cb8e730ad15", "node_type": "1", "metadata": {"page_label": "8", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "05aa0d513f045c09501bc1c6acc1ada1f5cccde7c0bb58eb9b669d4ba4b4ea7b"}, "3": {"node_id": "008368f5-ed54-4b4a-8e7e-f808ace630b7", "node_type": "1", "metadata": {"page_label": "8", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a7a87586ce64166ee85aa5f17a7b9258a209cef9776eefce570115032adee3c7"}}, "hash": "9bcaafea70f31abe07fed514eb4cca9e371a25e05ad6126cf9fe9a0d1cace3ca", "text": "different Bi-LSTMs for different node type for aggre-\ngating neighbor information. We re-implement this model\nin PyG following the authors\u2019 official code.\n\u2022Heterogeneous Graph Attention Networks (HAN) [ 23] de-\nsign hierarchical attentions to aggregate neighbor informa-\ntion via different meta paths. We re-implement this model\nin PyG following the authors\u2019 official code.\nIn addition, to systematically analyze the effectiveness of the\ntwo major components of HGT, i.e., Heterogeneous weight pa-\nrameterization (Heter) and Relative Temporal Encoding (RTE), we\nconduct an ablation study, but comparing with models that remove\nthese components. Specifically, we use \u2212Heter to denote models\nthat uses the same set of weights for all meta relations, and use\n\u2212RT E to denote models that doesn\u2019t include relative temporal en-\ncoding. By considering all the permutations, we have: HGT\u2212RT E\n\u2212Heter,\nHGT+RT E\n\u2212Heter, HGT\u2212RT E\n+Heterand HGT+RT E\n+Heter\u2021.\nWe use our HGSampling algorithm proposed in Section 4 for\nall baseline GNNs to handle the large-scale OAG graph. To avoid\ndata leakage, we remove out the links we aim to predict (e.g., the\nPaper-Field link as the label) from the sub-graph.\nInput Features. As we don\u2019t assume the feature of each node\ntype belongs to the same distribution, we are free to use the most\nappropriate features to represent each type of nodes. For each paper,\nwe use a pre-trained XLNet [ 24,25] to get the representation of each\nword in its title. We then average them weighted by each word\u2019s\nattention to get the title representation for each paper. The initial\nfeature of each author is then simply an average of his/her published\npapers\u2019 representations. For the field, venue, and institute nodes,\nwe use the metapath2vec model [ 3] to train their node embeddings\nby reflecting the heterogeneous network", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "008368f5-ed54-4b4a-8e7e-f808ace630b7": {"__data__": {"id_": "008368f5-ed54-4b4a-8e7e-f808ace630b7", "embedding": null, "metadata": {"page_label": "8", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using a pre-trained XLNet model in the Heterogeneous Graph Transformer (HGT) model for representing the title of each paper?\n2. How does the HGT model handle the heterogeneous network structures in the field, venue, and institute nodes?\n3. What is the role of the adaptation layer in the HGT model for making a fair comparison with homogeneous GNN baselines?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) model, which is a type of graph neural network (GNN) designed to handle heterogeneous graphs with multiple types of nodes. The model uses a combination of heterogeneous weight parameterization and relative temporal encoding to aggregate neighbor information and make predictions. The section also describes the HGSampling algorithm proposed to handle large-scale graphs and avoid data leakage. Input features for each type of node are discussed, including using pre-trained word embeddings for paper nodes and metapath2vec for field, venue, and institute nodes. The section concludes with an ablation study comparing the effectiveness of different components of the HGT model.", "section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) model, which is used for representing the title of each paper and handling heterogeneous network structures in the field, venue, and institute nodes. The model uses a pre-trained XLNet model to get the representation of each word in the title and an adaptation layer to make a fair comparison with homogeneous GNN baselines. The implementation details of the model are also discussed, including the use of 256 as the hidden dimension and 8 as the number of heads in the multi-head attention-based methods.", "excerpt_keywords": "1. Graph Neural Networks\n2. Heterogeneous Networks\n3. Node Embeddings\n4. Multi-Head Attention\n5. Adaptation Layer\n6. Field, Venue, and Institute Nodes\n7. Pre-trained XLNet\n8. Metapath2vec Model\n9. Homogeneous GNN Baselines\n10. Fairness and Positivity"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "70efd5d5-bcda-4698-a7e4-c0ba91bf514a", "node_type": "4", "metadata": {"page_label": "8", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "70254454ab988b1a5a2393bcda5e976341fb56875518f7a867588d109b8f7365"}, "2": {"node_id": "36bb2f32-dc7b-4109-94a8-ec50e301c6a7", "node_type": "1", "metadata": {"page_label": "8", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "9bcaafea70f31abe07fed514eb4cca9e371a25e05ad6126cf9fe9a0d1cace3ca"}}, "hash": "a7a87586ce64166ee85aa5f17a7b9258a209cef9776eefce570115032adee3c7", "text": "free to use the most\nappropriate features to represent each type of nodes. For each paper,\nwe use a pre-trained XLNet [ 24,25] to get the representation of each\nword in its title. We then average them weighted by each word\u2019s\nattention to get the title representation for each paper. The initial\nfeature of each author is then simply an average of his/her published\npapers\u2019 representations. For the field, venue, and institute nodes,\nwe use the metapath2vec model [ 3] to train their node embeddings\nby reflecting the heterogeneous network structures.\nThe homogeneous GNN baselines assume the node features be-\nlong to the same distribution, while our feature extraction doesn\u2019t\nfulfill this assumption. To make a fair comparison, we add an adap-\ntation layer between the input features and all used GNNs. This\nmodule simply conducts different linear projections for nodes of\ndifferent types. Such a procedure can be regarded to map hetero-\ngeneous data into the same distribution, which is also adopted in\nliterature [23, 27].\nImplementation Details. We use 256 as the hidden dimension\nthroughout the neural networks for all baselines. For all multi-head\nattention-based methods, we set the head number as 8. All GNNs\nkeep 3 layers so that the receptive fields of each network are exactly\n\u2021Unless other stated, HGT refers to HGT+RT E\n+Heter.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "e2dd3028-2df8-4f0c-9631-86451db2c111": {"__data__": {"id_": "e2dd3028-2df8-4f0c-9631-86451db2c111", "embedding": null, "metadata": {"page_label": "9", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the performance of Heterogeneous Graph Transformer in terms of NDCG and MRR for the CSPaper\u2013Field (L1) task compared to other GNN models such as GCN, RGCN, GAT, HetGNN, HAN, HGT-RT, and HeterHGT-RT?\n2. How does the batch time of Heterogeneous Graph Transformer compare to other GNN models for the CSPaper\u2013Field (L1) task, and what is the impact of batch time on the performance of these models?\n3. What is the difference in performance between Heterogeneous Graph Transformer and other GNN models such as GCN, RGCN, GAT, HetGNN, HAN, HGT-RT, and HeterHGT-RT for the Paper\u2013VenueNDCG task, and what factors contribute to this difference?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) model, which is used for representing the title of each paper and handling heterogeneous network structures in the field, venue, and institute nodes. The model uses a pre-trained XLNet model to get the representation of each word in the title and an adaptation layer to make a fair comparison with homogeneous GNN baselines. The implementation details of the model are also discussed, including the use of 256 as the hidden dimension and 8 as the number of heads in the multi-head attention-based methods.", "section_summary": "The section discusses the performance of Heterogeneous Graph Transformer (HGT) in terms of NDCG and MRR for the CSPaper\u2013Field (L1) task compared to other GNN models such as GCN, RGCN, GAT, HetGNN, HAN, HGT-RT, and HeterHGT-RT. It also compares the batch time of HGT to other GNN models for the CSPaper\u2013Field (L1) task and discusses the difference in performance between HGT and other GNN models for the Paper\u2013VenueNDCG task. The section also mentions the number of parameters and batch time of each model, as well as the factors that contribute to the difference in performance between HGT and other GNN models for the Paper\u2013VenueNDCG task.", "excerpt_keywords": "GNN Models, GCN, RGCN, GAT, HetGNN, HAN, HGT-RT, E-HeterHGT+RT, E-HeterHGT-RT, E-HeterHGT+RT, NDCG, MRR, Paper-Field, L1, Paper-Field, L2, Paper-Venue, NDCG, MRR, Paper-Venue."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "7ee672fa-2873-49ed-bc6c-25464d3d468c", "node_type": "4", "metadata": {"page_label": "9", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "bbedaa482fd33c1a4c7c8299516043ed80c0df6cebbea56f9472fcc012c8900c"}, "3": {"node_id": "67058179-72da-4f35-adfc-ccbaa8300641", "node_type": "1", "metadata": {"page_label": "9", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "29fd4686272b7bea7e8ced1e9a934ad32d8a19a24f3cc5ceec2e4d6899a654e7"}}, "hash": "39e0c15a3c93cad2fa6be3fc34e71dce10ef658d0634c553da6ee6473132aa19", "text": "Heterogeneous Graph Transformer WWW \u201920, April 20\u201324, 2020, Taipei, Taiwan\nGNN Models GCN [9] RGCN [14] GAT [22] HetGNN [27] HAN [23] HGT\u2212RT E\n\u2212HeterHGT+RT E\n\u2212HeterHGT\u2212RT E\n+HeterHGT+RT E\n+Heter\n#of Parameters 1.69M 8.80M 1.69M 8.41M 9.45M 3.12M 3.88M 7.44M 8.20M\nBatch Time 0.46s 1.24s 0.97s 1.35s 2.27s 1.11s 1.14s 1.48s 1.50s\nCSPaper\u2013Field ( L1)NDCG .608\u00b1.062 .603\u00b1.065 .622\u00b1.071 .612\u00b1.063 .618\u00b1.058 .662\u00b1.051 .689\u00b1.042 .705\u00b1.036 .718\u00b1.014\nMRR .679\u00b1.069 .683\u00b1.056 .694\u00b1.065 .689\u00b1.060 .691\u00b1.051 .751\u00b1.036 .779\u00b1.027 .799\u00b1.023 .823\u00b1.019\nPaper\u2013Field ( L2)NDCG .344\u00b1.021 .322\u00b1.053 .357\u00b1.058 .346\u00b1.071 .352\u00b1.051 .362\u00b1.048 .371\u00b1.043 .379\u00b1.047 .403\u00b1.041\nMRR .353\u00b1.053 .340\u00b1.061 .382\u00b1.057 .373\u00b1.051 .388\u00b1.065 .394\u00b1.072 .397\u00b1.064 .414\u00b1.076 .439\u00b1.078\nPaper\u2013VenueNDCG .406\u00b1.081 .412\u00b1.076 .437\u00b1.082 .431\u00b1.074 .449\u00b1.072 .456\u00b1.069 .461\u00b1.066 .468\u00b1.074 .473\u00b1.054\nMRR .215\u00b1.066", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "67058179-72da-4f35-adfc-ccbaa8300641": {"__data__": {"id_": "67058179-72da-4f35-adfc-ccbaa8300641", "embedding": null, "metadata": {"page_label": "9", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the performance of the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation in terms of NDCG and MRR?\n2. How does the disambiguation process affect the performance of the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation?\n3. How does the MedPaper\u2013Field (L1) and MedPaper\u2013Field (L2) metrics compare in terms of NDCG and MRR for the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation?", "prev_section_summary": "The section discusses the performance of Heterogeneous Graph Transformer (HGT) in terms of NDCG and MRR for the CSPaper\u2013Field (L1) task compared to other GNN models such as GCN, RGCN, GAT, HetGNN, HAN, HGT-RT, and HeterHGT-RT. It also compares the batch time of HGT to other GNN models for the CSPaper\u2013Field (L1) task and discusses the difference in performance between HGT and other GNN models for the Paper\u2013VenueNDCG task. The section also mentions the number of parameters and batch time of each model, as well as the factors that contribute to the difference in performance between HGT and other GNN models for the Paper\u2013VenueNDCG task.", "section_summary": "The section discusses the performance of a Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation in terms of NDCG and MRR. The disambiguation process is also discussed and its effect on the performance of the model is analyzed. Additionally, the MedPaper\u2013Field (L1) and MedPaper\u2013Field (L2) metrics are compared in terms of NDCG and MRR for the model. The section provides tables with the results of the experiments, including the performance metrics and the comparison of the metrics.", "excerpt_keywords": "Paper, Venue, NDCG, MRR, Author, Disambiguation, MedPaper\u2013Field ( L1), MedPaper\u2013Field ( L2), Paper\u2013Field ( L1), Paper\u2013Field ( L2), Paper\u2013VenueNDCG, Paper\u2013VenueMRR"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "7ee672fa-2873-49ed-bc6c-25464d3d468c", "node_type": "4", "metadata": {"page_label": "9", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "bbedaa482fd33c1a4c7c8299516043ed80c0df6cebbea56f9472fcc012c8900c"}, "2": {"node_id": "e2dd3028-2df8-4f0c-9631-86451db2c111", "node_type": "1", "metadata": {"page_label": "9", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "39e0c15a3c93cad2fa6be3fc34e71dce10ef658d0634c553da6ee6473132aa19"}, "3": {"node_id": "a21bd965-501e-42d3-b74b-c49995ad5f4f", "node_type": "1", "metadata": {"page_label": "9", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "17862337f3045e577fccae9bae95e0979a5a496e8a26f3b6cba1a5b2cc3eaa6a"}}, "hash": "29fd4686272b7bea7e8ced1e9a934ad32d8a19a24f3cc5ceec2e4d6899a654e7", "text": ".353\u00b1.053 .340\u00b1.061 .382\u00b1.057 .373\u00b1.051 .388\u00b1.065 .394\u00b1.072 .397\u00b1.064 .414\u00b1.076 .439\u00b1.078\nPaper\u2013VenueNDCG .406\u00b1.081 .412\u00b1.076 .437\u00b1.082 .431\u00b1.074 .449\u00b1.072 .456\u00b1.069 .461\u00b1.066 .468\u00b1.074 .473\u00b1.054\nMRR .215\u00b1.066 .216\u00b1.105 .239\u00b1.089 .245\u00b1.069 .254\u00b1.074 .258\u00b1.085 .265\u00b1.090 .275\u00b1.089 .288\u00b1.088\nAuthor\nDisambiguationNDCG .826\u00b1.039 .835\u00b1.042 .864\u00b1.051 .850\u00b1.056 .859\u00b1.053 .867\u00b1.048 .875\u00b1.046 .886\u00b1.048 .894\u00b1.034\nMRR .661\u00b1.045 .665\u00b1.054 .694\u00b1.052 .668\u00b1.061 .688\u00b1.049 .703\u00b1.036 .712\u00b1.032 .727\u00b1.038 .732\u00b1.038\nMedPaper\u2013Field ( L1)NDCG .560\u00b1.056 .571\u00b1.061 .584\u00b1.076 .598\u00b1.068 .607\u00b1.054 .654\u00b1.048 .667\u00b1.045 .683\u00b1.037 .709\u00b1.029\nMRR .465\u00b1.055 .470\u00b1.082 .493\u00b1.069 .509\u00b1.054 .575\u00b1.057 .620\u00b1.066 .642\u00b1.062 .659\u00b1.055 .688\u00b1.048\nPaper\u2013Field ( L2)NDCG .334\u00b1.035 .337\u00b1.051 .344\u00b1.063 .342\u00b1.048 .350\u00b1.059 .359\u00b1.053 .365\u00b1.047 .374\u00b1.050 .384\u00b1.046\nMRR .337\u00b1.061 .343\u00b1.063", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "a21bd965-501e-42d3-b74b-c49995ad5f4f": {"__data__": {"id_": "a21bd965-501e-42d3-b74b-c49995ad5f4f", "embedding": null, "metadata": {"page_label": "9", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the performance of the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation in terms of NDCG and MRR on the Paper\u2013Field and Paper\u2013Venue tasks?\n2. How does the disambiguation performance of the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation compare to other models on the Author task?\n3. What is the impact of the L1 and L2 regularization on the performance of the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation on the Paper\u2013Field and Paper\u2013Venue tasks?", "prev_section_summary": "The section discusses the performance of a Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation in terms of NDCG and MRR. The disambiguation process is also discussed and its effect on the performance of the model is analyzed. Additionally, the MedPaper\u2013Field (L1) and MedPaper\u2013Field (L2) metrics are compared in terms of NDCG and MRR for the model. The section provides tables with the results of the experiments, including the performance metrics and the comparison of the metrics.", "section_summary": "The section discusses the performance of a Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation on various tasks. The network is evaluated on the Paper\u2013Field and Paper\u2013Venue tasks for NDCG and MRR, and on the Author task for disambiguation performance. The results show that the network achieves high performance on all tasks, with NDCG and MRR values ranging from .334 to .718 and disambiguation MRR values ranging from .776 to .871. The section also discusses the impact of L1 and L2 regularization on the performance of the network on the Paper\u2013Field and Paper\u2013Venue tasks.", "excerpt_keywords": "Paper\u2013Field (L2)NDCG, Paper\u2013VenueNDCG, Author DisambiguationMRR, OAGPaper\u2013Field (L1)NDCG, MRR, NDCG, Paper\u2013Field, Paper\u2013Venue, Author, Disambiguation"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "7ee672fa-2873-49ed-bc6c-25464d3d468c", "node_type": "4", "metadata": {"page_label": "9", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "bbedaa482fd33c1a4c7c8299516043ed80c0df6cebbea56f9472fcc012c8900c"}, "2": {"node_id": "67058179-72da-4f35-adfc-ccbaa8300641", "node_type": "1", "metadata": {"page_label": "9", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "29fd4686272b7bea7e8ced1e9a934ad32d8a19a24f3cc5ceec2e4d6899a654e7"}, "3": {"node_id": "8054dd11-78a3-4777-b1ea-55132e226d1a", "node_type": "1", "metadata": {"page_label": "9", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "b3a11c8c6cbd5451f280fd44be3f990f14436ad513bd6a4701484fc73d672043"}}, "hash": "17862337f3045e577fccae9bae95e0979a5a496e8a26f3b6cba1a5b2cc3eaa6a", "text": ".465\u00b1.055 .470\u00b1.082 .493\u00b1.069 .509\u00b1.054 .575\u00b1.057 .620\u00b1.066 .642\u00b1.062 .659\u00b1.055 .688\u00b1.048\nPaper\u2013Field ( L2)NDCG .334\u00b1.035 .337\u00b1.051 .344\u00b1.063 .342\u00b1.048 .350\u00b1.059 .359\u00b1.053 .365\u00b1.047 .374\u00b1.050 .384\u00b1.046\nMRR .337\u00b1.061 .343\u00b1.063 .370\u00b1.058 .373\u00b1.061 .379\u00b1.052 .385\u00b1.071 .397\u00b1.069 .408\u00b1.071 .417\u00b1.074\nPaper\u2013VenueNDCG .377\u00b1.059 .383\u00b1.062 .388\u00b1.065 .412\u00b1.057 .416\u00b1.068 .421\u00b1.083 .432\u00b1.078 .446\u00b1.083 .445\u00b1.085\nMRR .211\u00b1.045 .217\u00b1.058 .244\u00b1.091 .259\u00b1.072 .271\u00b1.056 .277\u00b1.081 .282\u00b1.085 .288\u00b1.074 .291\u00b1.062\nAuthor\nDisambiguationMRR .776\u00b1.042 .779\u00b1.048 .828\u00b1.044 .824\u00b1.058 .834\u00b1.056 .838\u00b1.047 .844\u00b1.041 .864\u00b1.043 .871\u00b1.040\nNDCG .614\u00b1.051 .625\u00b1.049 .663\u00b1.046 .659\u00b1.061 .667\u00b1.053 .683\u00b1.055 .691\u00b1.046 .708\u00b1.041 .718\u00b1.043\nOAGPaper\u2013Field ( L1)NDCG .508\u00b1.141 .511\u00b1.128 .534\u00b1.103 .543\u00b1.084 .544\u00b1.096 .571\u00b1.089 .578\u00b1.086 .595\u00b1.089 .615\u00b1.084\nMRR .556\u00b1.136 .565\u00b1.105 .610\u00b1.096", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "8054dd11-78a3-4777-b1ea-55132e226d1a": {"__data__": {"id_": "8054dd11-78a3-4777-b1ea-55132e226d1a", "embedding": null, "metadata": {"page_label": "9", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the performance of the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation in terms of NDCG, MRR, and Author DisambiguationNDCG?\n2. How does the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation compare to other graph neural networks in terms of performance on the OAGPaper\u2013Field ( L1), Paper\u2013Field ( L2), Paper\u2013VenueNDCG, and Author DisambiguationNDCG metrics?\n3. What are the key features of the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation that contribute to its high performance on the NDCG, MRR, and Author DisambiguationNDCG metrics?", "prev_section_summary": "The section discusses the performance of a Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation on various tasks. The network is evaluated on the Paper\u2013Field and Paper\u2013Venue tasks for NDCG and MRR, and on the Author task for disambiguation performance. The results show that the network achieves high performance on all tasks, with NDCG and MRR values ranging from .334 to .718 and disambiguation MRR values ranging from .776 to .871. The section also discusses the impact of L1 and L2 regularization on the performance of the network on the Paper\u2013Field and Paper\u2013Venue tasks.", "section_summary": "The section discusses the performance of a Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation in terms of NDCG, MRR, and Author DisambiguationNDCG. The network is compared to other graph neural networks in terms of performance on the OAGPaper\u2013Field ( L1), Paper\u2013Field ( L2), Paper\u2013VenueNDCG, and Author DisambiguationNDCG metrics. The key features of the network that contribute to its high performance are not explicitly stated in the provided excerpt.", "excerpt_keywords": "NDCG, MRR, Paper-Field, Paper-Venue, Author, Disambiguation, OAGPaper, Field, Venue, Author"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "7ee672fa-2873-49ed-bc6c-25464d3d468c", "node_type": "4", "metadata": {"page_label": "9", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "bbedaa482fd33c1a4c7c8299516043ed80c0df6cebbea56f9472fcc012c8900c"}, "2": {"node_id": "a21bd965-501e-42d3-b74b-c49995ad5f4f", "node_type": "1", "metadata": {"page_label": "9", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "17862337f3045e577fccae9bae95e0979a5a496e8a26f3b6cba1a5b2cc3eaa6a"}, "3": {"node_id": "89e86816-c6c5-42d5-bc8b-ab254044ee5a", "node_type": "1", "metadata": {"page_label": "9", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "fa9362cdc2271a6e68e5e077e2d1fa84df56b0224af487a33287547ad92b723d"}}, "hash": "b3a11c8c6cbd5451f280fd44be3f990f14436ad513bd6a4701484fc73d672043", "text": ".625\u00b1.049 .663\u00b1.046 .659\u00b1.061 .667\u00b1.053 .683\u00b1.055 .691\u00b1.046 .708\u00b1.041 .718\u00b1.043\nOAGPaper\u2013Field ( L1)NDCG .508\u00b1.141 .511\u00b1.128 .534\u00b1.103 .543\u00b1.084 .544\u00b1.096 .571\u00b1.089 .578\u00b1.086 .595\u00b1.089 .615\u00b1.084\nMRR .556\u00b1.136 .565\u00b1.105 .610\u00b1.096 .616\u00b1.076 .622\u00b1.092 .649\u00b1.081 .657\u00b1.078 .675\u00b1.082 .702\u00b1.081\nPaper\u2013Field ( L2)NDCG .318\u00b1.074 .328\u00b1.046 .339\u00b1.049 .336\u00b1.062 .342\u00b1.051 .350\u00b1.045 .354\u00b1.046 .358\u00b1.052 .367\u00b1.048\nMRR .322\u00b1.067 .332\u00b1.052 .348\u00b1.045 .350\u00b1.053 .358\u00b1.049 .362\u00b1.057 .369\u00b1.058 .371\u00b1.064 .378\u00b1.071\nPaper\u2013VenueNDCG .302\u00b1.066 .313\u00b1.051 .317\u00b1.057 .309\u00b1.071 .327\u00b1.062 .334\u00b1.058 .341\u00b1.059 .353\u00b1.064 .355\u00b1.062\nMRR .194\u00b1.070 .193\u00b1.047 .196\u00b1.052 .192\u00b1.059 .214\u00b1.067 .229\u00b1.061 .233\u00b1.060 .243\u00b1.048 .247\u00b1.061\nAuthor\nDisambiguationNDCG .738\u00b1.042 .755\u00b1.048 .797\u00b1.044 .803\u00b1.058 .821\u00b1.056 .835\u00b1.043 .841\u00b1.041 .847\u00b1.043 .852\u00b1.048\nMRR .612\u00b1.064 .619\u00b1.057 .645\u00b1.063 .649\u00b1.052", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "89e86816-c6c5-42d5-bc8b-ab254044ee5a": {"__data__": {"id_": "89e86816-c6c5-42d5-bc8b-ab254044ee5a", "embedding": null, "metadata": {"page_label": "9", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the proposed method for corporate relative valuation and how does it compare to other baselines in terms of NDCG and MRR?\n2. What is the significance of the proposed method in terms of its performance gains over baselines for the Paper\u2013Field (L1) classification task on OAG?\n3. How does the proposed method perform in comparison to other baselines on the CS, Med, and OAG datasets in terms of relative NDCG improvements?", "prev_section_summary": "The section discusses the performance of a Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation in terms of NDCG, MRR, and Author DisambiguationNDCG. The network is compared to other graph neural networks in terms of performance on the OAGPaper\u2013Field ( L1), Paper\u2013Field ( L2), Paper\u2013VenueNDCG, and Author DisambiguationNDCG metrics. The key features of the network that contribute to its high performance are not explicitly stated in the provided excerpt.", "section_summary": "The section discusses a proposed method for corporate relative valuation using a heterogeneous graph neural network (HGT) and compares its performance to other baselines on three datasets. The proposed method uses a multi-modal graph neural network architecture that incorporates both textual and visual information. The experimental results show that the proposed method outperforms all baselines in terms of both NDCG and MRR for all tasks on all datasets. The average relative NDCG improvements of HGT on the CS, Med, and OAG datasets are 11%, 10%, and 8%, respectively. The proposed method also outperforms GCN, GAT, RGCN, HetGNN, and HAN by an average of 11%.", "excerpt_keywords": "1. Heterogeneous Graph Transformer (HGT)\n2. Graph Neural Network (GNN)\n3. AdamW optimizer\n4. Cosine Annealing Learning Rate Scheduler\n5. Disambiguation\n6. NDCG\n7. MRR\n8. Paper\u2013Field (L1) classification task\n9. OAG dataset\n10. Experimental results"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "7ee672fa-2873-49ed-bc6c-25464d3d468c", "node_type": "4", "metadata": {"page_label": "9", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "bbedaa482fd33c1a4c7c8299516043ed80c0df6cebbea56f9472fcc012c8900c"}, "2": {"node_id": "8054dd11-78a3-4777-b1ea-55132e226d1a", "node_type": "1", "metadata": {"page_label": "9", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "b3a11c8c6cbd5451f280fd44be3f990f14436ad513bd6a4701484fc73d672043"}, "3": {"node_id": "ceab7ae5-1fbe-4481-8356-475bf9774c29", "node_type": "1", "metadata": {"page_label": "9", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "99dbeb532efa045df545242e5e1c9e39fb73f7cf00a35d0041636d93704903a7"}}, "hash": "fa9362cdc2271a6e68e5e077e2d1fa84df56b0224af487a33287547ad92b723d", "text": ".196\u00b1.052 .192\u00b1.059 .214\u00b1.067 .229\u00b1.061 .233\u00b1.060 .243\u00b1.048 .247\u00b1.061\nAuthor\nDisambiguationNDCG .738\u00b1.042 .755\u00b1.048 .797\u00b1.044 .803\u00b1.058 .821\u00b1.056 .835\u00b1.043 .841\u00b1.041 .847\u00b1.043 .852\u00b1.048\nMRR .612\u00b1.064 .619\u00b1.057 .645\u00b1.063 .649\u00b1.052 .660\u00b1.049 .668\u00b1.059 .674\u00b1.058 .683\u00b1.066 .688\u00b1.054\nTable 2: Experimental results of different methods over the three datasets.\nthe same. All baselines are optimized via the AdamW optimizer [ 13]\nwith the Cosine Annealing Learning Rate Scheduler [ 12]. For each\nmodel, we train it for 200 epochs and select the one with the lowest\nvalidation loss as the reported model. We use the default parameters\nused in GNN literature and donot tune hyper-parameters.\n5.3 Experimental Results\nWe summarize the experimental results of the proposed model\nand baselines on three datasets in Table 2. All experiments for the\nfour tasks are evaluated in terms of NDCG and MRR.\nThe results show that in terms of both metrics, the proposed\nHGT significantly and consistently outperforms all baselines for\nall tasks on all datasets. Take, for example, the Paper\u2013Field ( L1)\nclassification task on OAG, HGT achieves relative performance\ngains over baselines by 15\u201319% in terms of NDCG and 18\u201321% interms of MRR (i.e., the performance gap divided by the baseline\nperformance). When compared to HAN\u2014the best baseline for most\nof the cases, the average relative NDCG improvements of HGT on\nthe CS, Med and OAG datasets are 11 %, 10%and 8 %, respectively.\nOverall, we observe that on average, HGT outperforms GCN,\nGAT, RGCN, HetGNN, and HAN by", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "ceab7ae5-1fbe-4481-8356-475bf9774c29": {"__data__": {"id_": "ceab7ae5-1fbe-4481-8356-475bf9774c29", "embedding": null, "metadata": {"page_label": "9", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Graph Transformer (HGT) and how does it improve the performance of graph neural networks?\n2. How does HGT compare to other heterogeneous graph neural network baselines in terms of performance and resource consumption?\n3. What are the specific advantages of modeling heterogeneous edges according to their meta relation schema in HGT?", "prev_section_summary": "The section discusses a proposed method for corporate relative valuation using a heterogeneous graph neural network (HGT) and compares its performance to other baselines on three datasets. The proposed method uses a multi-modal graph neural network architecture that incorporates both textual and visual information. The experimental results show that the proposed method outperforms all baselines in terms of both NDCG and MRR for all tasks on all datasets. The average relative NDCG improvements of HGT on the CS, Med, and OAG datasets are 11%, 10%, and 8%, respectively. The proposed method also outperforms GCN, GAT, RGCN, HetGNN, and HAN by an average of 11%.", "section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT), a graph neural network that improves the performance of graph neural networks by modeling heterogeneous edges according to their meta relation schema. HGT achieves relative performance gains of 15-19% in terms of NDCG and 18-21% in terms of MRR compared to other heterogeneous graph neural network baselines. HGT outperforms GCN, GAT, RGCN, HetGNN, and HAN by 20% for the four tasks on all three large-scale datasets, while having fewer parameters and comparable batch time.", "excerpt_keywords": "1. Heterogeneous graph neural networks, 2. Relative performance gains, 3. Neural diversity, 4. Meta relation schema, 5. Generalization, 6. Resource consumption, 7. Neural diversity, 8. Fairness and positivity, 9. Utmost utility, 10. Prejudiced or negative content."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "7ee672fa-2873-49ed-bc6c-25464d3d468c", "node_type": "4", "metadata": {"page_label": "9", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "bbedaa482fd33c1a4c7c8299516043ed80c0df6cebbea56f9472fcc012c8900c"}, "2": {"node_id": "89e86816-c6c5-42d5-bc8b-ab254044ee5a", "node_type": "1", "metadata": {"page_label": "9", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "fa9362cdc2271a6e68e5e077e2d1fa84df56b0224af487a33287547ad92b723d"}}, "hash": "99dbeb532efa045df545242e5e1c9e39fb73f7cf00a35d0041636d93704903a7", "text": "HGT achieves relative performance\ngains over baselines by 15\u201319% in terms of NDCG and 18\u201321% interms of MRR (i.e., the performance gap divided by the baseline\nperformance). When compared to HAN\u2014the best baseline for most\nof the cases, the average relative NDCG improvements of HGT on\nthe CS, Med and OAG datasets are 11 %, 10%and 8 %, respectively.\nOverall, we observe that on average, HGT outperforms GCN,\nGAT, RGCN, HetGNN, and HAN by 20% for the four tasks on all\nthree large-scale datasets. Moreover, HGT has fewer parameters\nand comparable batch time than all the heterogeneous graph neu-\nral network baselines, including RGCN, HetGNN, and HAN. This\nsuggests that by modeling heterogeneous edges according to their\nmeta relation schema, we are able to have better generalization\nwith fewer resource consumption.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "cbfd04f2-7520-418b-bd69-a2200a9c66ed": {"__data__": {"id_": "cbfd04f2-7520-418b-bd69-a2200a9c66ed", "embedding": null, "metadata": {"page_label": "10", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the significance of using Relative Temporal Encoding in Heterogeneous Graph Transformer (HGT) for capturing graph dynamics?\n2. How does the removal of heterogeneous weight parameterization and Relative Temporal Encoding affect the performance of HGT?\n3. Can using meta relations to parameterize weight matrices achieve competitive performance with fewer resources compared to a baseline that keeps a unique weight matrix for each relation?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT), a graph neural network that improves the performance of graph neural networks by modeling heterogeneous edges according to their meta relation schema. HGT achieves relative performance gains of 15-19% in terms of NDCG and 18-21% in terms of MRR compared to other heterogeneous graph neural network baselines. HGT outperforms GCN, GAT, RGCN, HetGNN, and HAN by 20% for the four tasks on all three large-scale datasets, while having fewer parameters and comparable batch time.", "section_summary": "The section discusses the significance of using Relative Temporal Encoding in Heterogeneous Graph Transformer (HGT) for capturing graph dynamics. An ablation study is conducted to analyze the effects of removing heterogeneous weight parameterization and Relative Temporal Encoding on the performance of HGT. The results show that parameterizing with meta relations and using Relative Temporal Encoding are significant for achieving competitive performance with fewer resources. A case study is also conducted to evaluate how Relative Temporal Encoding can help HGT to capture graph dynamics by showing the evolution of conference topics.", "excerpt_keywords": "Heterogeneous weight parameterization, Relative Temporal Encoding, ablation study, graph dynamics, conference topic, evolution, WWW, KDD, NeurIPS, timestamps, sub-graphs, representations, euclidean distances, similarity, relationships, change, keywords."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "29d3becc-38ca-47cb-9332-674e4b7909a6", "node_type": "4", "metadata": {"page_label": "10", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "f8e8c420fc0e4c1b5ef6663c7468203d1e2a7c84037324476427e0427a42a4ea"}, "3": {"node_id": "ec2fabc6-1aaf-4704-85b0-f81e974ffb56", "node_type": "1", "metadata": {"page_label": "10", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "bebf9c8a405f77ababb75fb0b31d69e96340d6d4ce837eda990cd2029dd79b45"}}, "hash": "8e2cefc001cc15dce78e78c2a04e4a67982849f9a4b37a98a633e9cc7021a463", "text": "WWW \u201920, April 20\u201324, 2020, Taipei, Taiwan Ziniu Hu, Yuxiao Dong, Kuansan Wang, and Yizhou Sun\nAblation Study. The core component in HGT are the heteroge-\nneous weight parameterization (Heter) and Relative Temporal En-\ncoding (RTE). To further analyze their effects, we conduct an abla-\ntion study by removing them from HGT. Specifically, the model that\nremoves heterogeneous weight parameterization, i.e., HGT+RT E\n\u2212Heter,\ndrops 4% of performance compared with the full model HGT+RT E\n+Heter.\nBy removing RTE (i.e., HGT\u2212RT E\n+Heter), the performance has a 2% drop.\nThe ablation study shows the significance of parameterizing with\nmeta relations and using Relative Temporal Encoding.\nIn addition, we also try to implement a baseline that keeps a\nunique weight matrix for each relation. However, such a baseline\ncontains too many parameters so that our experimental setting\ndoesn\u2019t have enough GPU memory to optimize it. This also indicates\nthat using the meta relation to parameterize weight matrices can\nachieve competitive performance with fewer resources.\n5.4 Case Study\nTo further evaluate how Relative Temporal Encoding (RTE) can\nhelp HGT to capture graph dynamics, we conduct a case study\nshowing the evolution of conference topic. We select 100 confer-\nences in computer science with the highest citations, assign them\nthree different timestamps, i.e., 2000, 2010 and 2020, and construct\nsub-graphs initialized by them. Using a trained HGT, we can get\nthe representations for these conferences, with which we can calcu-\nlate the euclidean distances between them. We select WWW, KDD,\nand NeurIPS as illustration. For each of them, we pick the top-5\nmost similar conferences (i.e., the one with the smallest euclidean\ndistance) to show how the conference\u2019s topics evolve over time.\nAs shown in Table 3, these venues\u2019 relationships have changed\nfrom 2000", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "ec2fabc6-1aaf-4704-85b0-f81e974ffb56": {"__data__": {"id_": "ec2fabc6-1aaf-4704-85b0-f81e974ffb56", "embedding": null, "metadata": {"page_label": "10", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. How does the Heterogeneous Graph Transformer model help capture the temporal evolution of heterogeneous academic graphs?\n2. Can you explain how the meta relation schema benefits the heterogeneous message passing process in the Heterogeneous Graph Transformer model?\n3. How does the HGT model help identify the most similar conferences to a given conference based on their topics and relationships?", "prev_section_summary": "The section discusses the significance of using Relative Temporal Encoding in Heterogeneous Graph Transformer (HGT) for capturing graph dynamics. An ablation study is conducted to analyze the effects of removing heterogeneous weight parameterization and Relative Temporal Encoding on the performance of HGT. The results show that parameterizing with meta relations and using Relative Temporal Encoding are significant for achieving competitive performance with fewer resources. A case study is also conducted to evaluate how Relative Temporal Encoding can help HGT to capture graph dynamics by showing the evolution of conference topics.", "section_summary": "The section describes the use of a Heterogeneous Graph Transformer (HGT) model to capture the temporal evolution of heterogeneous academic graphs. The model is trained on sub-graphs initialized by specific conferences and their relationships, and can be used to calculate the similarity between conferences based on their topics and relationships. The section also includes a case study that demonstrates the model's ability to capture the difference brought by new conferences. Additionally, the section discusses the use of a meta relation schema to benefit the heterogeneous message passing process in the HGT model, and includes a visualization of the meta relation attention hierarchy tree.", "excerpt_keywords": "1. Heterogeneous Graphs\n2. Temporal Evolution\n3. Meta Relation Attention\n4. Heterogeneous Message Passing\n5. Graph Neural Networks\n6. Academic Graphs\n7. Conference Relationships\n8. Topic Evolution\n9. Database Conferences\n10. Networking Conferences"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "29d3becc-38ca-47cb-9332-674e4b7909a6", "node_type": "4", "metadata": {"page_label": "10", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "f8e8c420fc0e4c1b5ef6663c7468203d1e2a7c84037324476427e0427a42a4ea"}, "2": {"node_id": "cbfd04f2-7520-418b-bd69-a2200a9c66ed", "node_type": "1", "metadata": {"page_label": "10", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "8e2cefc001cc15dce78e78c2a04e4a67982849f9a4b37a98a633e9cc7021a463"}, "3": {"node_id": "9ec2d462-902a-4147-98f1-ce3d1c978d93", "node_type": "1", "metadata": {"page_label": "10", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "5430ad2917d6e4349804f6962b2792e8598333b93f5bf77d5eacdc3b4026ae6c"}}, "hash": "bebf9c8a405f77ababb75fb0b31d69e96340d6d4ce837eda990cd2029dd79b45", "text": "2010 and 2020, and construct\nsub-graphs initialized by them. Using a trained HGT, we can get\nthe representations for these conferences, with which we can calcu-\nlate the euclidean distances between them. We select WWW, KDD,\nand NeurIPS as illustration. For each of them, we pick the top-5\nmost similar conferences (i.e., the one with the smallest euclidean\ndistance) to show how the conference\u2019s topics evolve over time.\nAs shown in Table 3, these venues\u2019 relationships have changed\nfrom 2000 to 2020. For example, WWW in 2000 was more related to\nsome database conferences, i.e., SIGMOD and VLDB, and some net-\nworking conferences, i.e., NSDI and GLOBECOM. However, WWW\nin 2020 would become more related to some data mining and infor-\nmation retrieval conferences (KDD, SIGIR, and WSDM), in addition\nto SIGMOD and GLOBECOM. Also, KDD in 2000 was more related\nto traditional database and data mining venues, while in 2020 it\nwill tend to correlate with a variety of topics, i.e. machine learning\n(NeurIPS), database (SIGMOD), Web (WWW), AI (AAAI), and NLP\n(EMNLP). Additionally, our HGT model can capture the difference\nbrought by new conferences. For example, NeurIPS in 2020 would\nrelate with ICLR, which is a newly organized deep learning confer-\nence. This case study shows that the relative temporal encoding can\nhelp capture the temporal evolution of the heterogeneous academic\ngraphs.\n5.5 Visualize Meta Relation Attention\nTo illustrate how the incorporated meta relation schema can\nbenefit the heterogeneous message passing process, we pick the\nschema that has the largest attention value in each of the first\ntwo HGT layers and plot the meta relation attention hierarchy\ntree in Figure 5. For example, to calculate a paper\u2019s representation,\n\u27e8Paper, is_published _at, Venue, is_published _at\u22121,", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "9ec2d462-902a-4147-98f1-ce3d1c978d93": {"__data__": {"id_": "9ec2d462-902a-4147-98f1-ce3d1c978d93", "embedding": null, "metadata": {"page_label": "10", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. How does the relative temporal encoding help capture the temporal evolution of heterogeneous academic graphs?\n2. What is the process of visualizing meta relation attention in the Heterogeneous Graph Transformer model?\n3. Can you explain how the importance of meta paths is automatically learned from the data without manual design in the Heterogeneous Graph Transformer model?", "prev_section_summary": "The section describes the use of a Heterogeneous Graph Transformer (HGT) model to capture the temporal evolution of heterogeneous academic graphs. The model is trained on sub-graphs initialized by specific conferences and their relationships, and can be used to calculate the similarity between conferences based on their topics and relationships. The section also includes a case study that demonstrates the model's ability to capture the difference brought by new conferences. Additionally, the section discusses the use of a meta relation schema to benefit the heterogeneous message passing process in the HGT model, and includes a visualization of the meta relation attention hierarchy tree.", "section_summary": "The section discusses the use of the Heterogeneous Graph Transformer (HGT) model for corporate relative valuation. The model uses relative temporal encoding to capture the temporal evolution of heterogeneous academic graphs. The section also explains how the importance of meta paths is automatically learned from the data without manual design in the HGT model. The section provides examples of calculating paper and author node representations using the learned meta paths. The section also includes a table and figure showing the temporal evolution of conference similarity and the hierarchy of the learned meta relation attention.", "excerpt_keywords": "1. Heterogeneous academic graphs\n2. Temporal evolution\n3. Meta relation attention\n4. Meta relation schema\n5. Meta paths\n6. Venue\n7. Time\n8. Topic\n9. Similarity\n10. Keyword extraction"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "29d3becc-38ca-47cb-9332-674e4b7909a6", "node_type": "4", "metadata": {"page_label": "10", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "f8e8c420fc0e4c1b5ef6663c7468203d1e2a7c84037324476427e0427a42a4ea"}, "2": {"node_id": "ec2fabc6-1aaf-4704-85b0-f81e974ffb56", "node_type": "1", "metadata": {"page_label": "10", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "bebf9c8a405f77ababb75fb0b31d69e96340d6d4ce837eda990cd2029dd79b45"}, "3": {"node_id": "68b2b3e6-15b0-4bfc-b376-00288ff26c68", "node_type": "1", "metadata": {"page_label": "10", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "7b9c3bb7b19c99309614af9b5130cc88bbd97e2e92544909217178ae4a817137"}}, "hash": "5430ad2917d6e4349804f6962b2792e8598333b93f5bf77d5eacdc3b4026ae6c", "text": "This case study shows that the relative temporal encoding can\nhelp capture the temporal evolution of the heterogeneous academic\ngraphs.\n5.5 Visualize Meta Relation Attention\nTo illustrate how the incorporated meta relation schema can\nbenefit the heterogeneous message passing process, we pick the\nschema that has the largest attention value in each of the first\ntwo HGT layers and plot the meta relation attention hierarchy\ntree in Figure 5. For example, to calculate a paper\u2019s representation,\n\u27e8Paper, is_published _at, Venue, is_published _at\u22121, Paper\u27e9,\u27e8Paper,\nhas_L2_f ield _of, Field, has_L5_f ield _of\u22121, Paper\u27e9, and\u27e8Institute,\nis_af f iliated _with\u22121, Author, is_f irst _author _of, Paper\u27e9are the\nthree most important meta relation sequences, which can be re-\ngarded as meta paths PVP, PFP, and IAP, respectively. Note thatVenue Time Top \u22125 Most Similar Venues\nWWW2000 SIGMOD, VLDB, NSDI, GLOBECOM, SIGIR\n2010 GLOBECOM, KDD, CIKM, SIGIR, SIGMOD\n2020 KDD, GLOBECOM, SIGIR, WSDM, SIGMOD\nKDD2000 SIGMOD, ICDE, ICDM, CIKM, VLDB\n2010 ICDE, WWW, NeurIPS, SIGMOD, ICML\n2020 NeurIPS, SIGMOD, WWW, AAAI, EMNLP\nNeurIPS2000 ICCV, ICML, ECCV, AAAI, CVPR\n2010 ICML, CVPR, ACL, KDD, AAAI\n2020 ICML, CVPR, ICLR, ICCV, ACL\nTable 3: Temporal Evolution of Conference Similarity.\nFigure 5: Hierarchy of the learned meta relation attention.\nthese meta paths and their importance are automatically learned\nfrom the data without manual design. Another example of calcu-\nlating an author node\u2019s representation is shown on the", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "68b2b3e6-15b0-4bfc-b376-00288ff26c68": {"__data__": {"id_": "68b2b3e6-15b0-4bfc-b376-00288ff26c68", "embedding": null, "metadata": {"page_label": "10", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the proposed architecture for modeling heterogeneous and dynamic graphs in the paper \"Heterogeneous Graph Transformer\"?\n2. How does the relative temporal encoding (RTE) technique incorporate temporal information in the proposed architecture?\n3. What is the heterogeneous Mini-Batch graph sampling algorithm (HGSampling) and how does it facilitate efficient and scalable training of the proposed architecture on Web-scale data?", "prev_section_summary": "The section discusses the use of the Heterogeneous Graph Transformer (HGT) model for corporate relative valuation. The model uses relative temporal encoding to capture the temporal evolution of heterogeneous academic graphs. The section also explains how the importance of meta paths is automatically learned from the data without manual design in the HGT model. The section provides examples of calculating paper and author node representations using the learned meta paths. The section also includes a table and figure showing the temporal evolution of conference similarity and the hierarchy of the learned meta relation attention.", "section_summary": "The section discusses the proposed architecture for modeling heterogeneous and dynamic graphs in the paper \"Heterogeneous Graph Transformer\". The architecture incorporates temporal information using the relative temporal encoding (RTE) technique and facilitates efficient and scalable training on Web-scale data using the heterogeneous Mini-Batch graph sampling algorithm (HGSampling). The paper also shows that the proposed architecture can capture both heterogeneity and outperforms state-of-the-art GNN baselines on various downstream tasks. The authors also mention future work to explore whether the model can generate heterogeneous graphs and pre-train it to benefit tasks with scarce labels.", "excerpt_keywords": "Heterogeneous Graph Transformer, Web-scale heterogeneous and dynamic graphs, meta relation, temporal encoding, relative temporal encoding, heterogeneous Mini-Batch graph sampling algorithm, Open Academic Graph, downstream tasks, graph dynamics, heterogeneity."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "29d3becc-38ca-47cb-9332-674e4b7909a6", "node_type": "4", "metadata": {"page_label": "10", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "f8e8c420fc0e4c1b5ef6663c7468203d1e2a7c84037324476427e0427a42a4ea"}, "2": {"node_id": "9ec2d462-902a-4147-98f1-ce3d1c978d93", "node_type": "1", "metadata": {"page_label": "10", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "5430ad2917d6e4349804f6962b2792e8598333b93f5bf77d5eacdc3b4026ae6c"}, "3": {"node_id": "f9877c88-0e08-4067-b0ee-40dac187596f", "node_type": "1", "metadata": {"page_label": "10", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "fb4873731585981ccb15e2af00b1c7f621baecf64cfd561157f788104454fd9c"}}, "hash": "7b9c3bb7b19c99309614af9b5130cc88bbd97e2e92544909217178ae4a817137", "text": "SIGMOD, WWW, AAAI, EMNLP\nNeurIPS2000 ICCV, ICML, ECCV, AAAI, CVPR\n2010 ICML, CVPR, ACL, KDD, AAAI\n2020 ICML, CVPR, ICLR, ICCV, ACL\nTable 3: Temporal Evolution of Conference Similarity.\nFigure 5: Hierarchy of the learned meta relation attention.\nthese meta paths and their importance are automatically learned\nfrom the data without manual design. Another example of calcu-\nlating an author node\u2019s representation is shown on the right. Such\nvisualization demonstrates that Heterogeneous Graph Transformer\nis capable of implicitly learning to construct important meta paths\nfor specific downstream tasks, without manual customization.\n6 CONCLUSION\nIn this paper, we propose the Heterogeneous Graph Transformer\n(HGT) architecture for modeling Web-scale heterogeneous and dy-\nnamic graphs. To model heterogeneity, we use the meta relation\n\u27e8\u03c4(s),\u03d5(e),\u03c4(t)\u27e9to decompose the interaction and transform matri-\nces, enabling the model to have the similar modeling capacity with\nfewer resources. To capture graph dynamics, we present the relative\ntemporal encoding (RTE) technique to incorporate temporal infor-\nmation using limited computational resources. To conduct efficient\nand scalable training of HGT on Web-scale data, we design the het-\nerogeneous Mini-Batch graph sampling algorithm\u2014HGSampling.\nWe conduct comprehensive experiments on the Open Academic\nGraph, and show that the proposed HGT model can capture both\nheterogeneity and outperforms all the state-of-the-art GNN base-\nlines on various downstream tasks.\nIn the future, we will explore whether HGT is able to generate\nheterogeneous graphs, e.g., predict new papers and their titles, and\nwhether we can pre-train HGT to benefit tasks with scarce labels.\nAcknowledgements. We would like to thank Xiaodong Liu for\nhelpful discussions. This work is partially supported by NSF III-\n1705169, NSF CAREER Award 1741634, NSF#1937599, Okawa", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "f9877c88-0e08-4067-b0ee-40dac187596f": {"__data__": {"id_": "f9877c88-0e08-4067-b0ee-40dac187596f", "embedding": null, "metadata": {"page_label": "10", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Graph Transformer (HGT) and how does it differ from other Graph Neural Network (GNN) base-lines?\n2. What are the potential future applications of HGT and how can it benefit tasks with scarce labels?\n3. Who were the key contributors to the development of HGT and what funding sources supported the research?", "prev_section_summary": "The section discusses the proposed architecture for modeling heterogeneous and dynamic graphs in the paper \"Heterogeneous Graph Transformer\". The architecture incorporates temporal information using the relative temporal encoding (RTE) technique and facilitates efficient and scalable training on Web-scale data using the heterogeneous Mini-Batch graph sampling algorithm (HGSampling). The paper also shows that the proposed architecture can capture both heterogeneity and outperforms state-of-the-art GNN baselines on various downstream tasks. The authors also mention future work to explore whether the model can generate heterogeneous graphs and pre-train it to benefit tasks with scarce labels.", "section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT), a type of Graph Neural Network (GNN) that is designed to handle heterogeneous graphs with multiple types of nodes and edges. HGT outperforms other GNN base-lines on various downstream tasks and has the potential to generate heterogeneous graphs, such as predicting new papers and their titles. The development of HGT was supported by funding sources including the National Science Foundation (NSF), Okawa Foundation, and Amazon Research Award. The key contributors to the development of HGT include Xiaodong Liu.", "excerpt_keywords": "1. Heterogeneous Graphs, 2. Graph Neural Networks (GNN), 3. Downstream Tasks, 4. State-of-the-Art, 5. Pre-Training, 6. Positive Content, 7. Fairness, 8. Utility, 9. Negative Content, 10. Ethical Considerations"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "29d3becc-38ca-47cb-9332-674e4b7909a6", "node_type": "4", "metadata": {"page_label": "10", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "f8e8c420fc0e4c1b5ef6663c7468203d1e2a7c84037324476427e0427a42a4ea"}, "2": {"node_id": "68b2b3e6-15b0-4bfc-b376-00288ff26c68", "node_type": "1", "metadata": {"page_label": "10", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "7b9c3bb7b19c99309614af9b5130cc88bbd97e2e92544909217178ae4a817137"}}, "hash": "fb4873731585981ccb15e2af00b1c7f621baecf64cfd561157f788104454fd9c", "text": "capture both\nheterogeneity and outperforms all the state-of-the-art GNN base-\nlines on various downstream tasks.\nIn the future, we will explore whether HGT is able to generate\nheterogeneous graphs, e.g., predict new papers and their titles, and\nwhether we can pre-train HGT to benefit tasks with scarce labels.\nAcknowledgements. We would like to thank Xiaodong Liu for\nhelpful discussions. This work is partially supported by NSF III-\n1705169, NSF CAREER Award 1741634, NSF#1937599, Okawa Foun-\ndation Grant, and Amazon Research Award.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "8a448c50-08b1-47f8-853b-919396983891": {"__data__": {"id_": "8a448c50-08b1-47f8-853b-919396983891", "embedding": null, "metadata": {"page_label": "11", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Graph Transformer paper presented at the WWW \u201920 conference?\n2. How does the FastGCN algorithm improve the efficiency of learning with Graph Convolutional Networks?\n3. What is the significance of the Stochastic Training of Graph Convolu-tional Networks with Variance Reduction technique in the field of graph neural networks?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT), a type of Graph Neural Network (GNN) that is designed to handle heterogeneous graphs with multiple types of nodes and edges. HGT outperforms other GNN base-lines on various downstream tasks and has the potential to generate heterogeneous graphs, such as predicting new papers and their titles. The development of HGT was supported by funding sources including the National Science Foundation (NSF), Okawa Foundation, and Amazon Research Award. The key contributors to the development of HGT include Xiaodong Liu.", "section_summary": "The section discusses the Heterogeneous Graph Transformer paper presented at the WWW '20 conference, which focuses on developing a multi-modal graph neural network for corporate relative valuation. The paper also mentions the FastGCN algorithm, which improves the efficiency of learning with Graph Convolutional Networks, and the Stochastic Training of Graph Convolu-tional Networks with Variance Reduction technique, which is significant in the field of graph neural networks. Additionally, the section references several other papers related to graph neural networks and representation learning.", "excerpt_keywords": "Graph convolutional networks, importance sampling, stochastic training, heterogeneous networks, metapath2vec, globalization, scientific collaborations, citations, innovations, neural message passing, quantum chemistry, inductive representation learning, deep residual learning, image recognition, semi-supervised classification."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "f3ccaf27-46f6-4aaa-b10d-195f2156f4b6", "node_type": "4", "metadata": {"page_label": "11", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a8b06d28c38dad7b3eeb48534f0f6dd917aa3349a031c7bf425429fd635407d3"}, "3": {"node_id": "96dc6fbf-2f14-4e07-ac4c-0f3385813d2b", "node_type": "1", "metadata": {"page_label": "11", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "7409ba8a8ce9e2c430770b45f2be1640f1e68830657d69f8bbd4a9ab86fb72c1"}}, "hash": "4193f4744201a0499d6bed904b345b1213475f8f2acab1f748ccc55fa4fa355d", "text": "Heterogeneous Graph Transformer WWW \u201920, April 20\u201324, 2020, Taipei, Taiwan\nREFERENCES\n[1]Jie Chen, Tengfei Ma, and Cao Xiao. 2018. FastGCN: Fast Learning with Graph\nConvolutional Networks via Importance Sampling. In ICLR\u201918 .\n[2]Jianfei Chen, Jun Zhu, and Le Song. 2018. Stochastic Training of Graph Convolu-\ntional Networks with Variance Reduction. In ICML . 941\u2013949.\n[3]Yuxiao Dong, Nitesh V Chawla, and Ananthram Swami. 2017. metapath2vec:\nScalable Representation Learning for Heterogeneous Networks. In KDD \u201917 .\n[4]Yuxiao Dong, Hao Ma, Zhihong Shen, and Kuansan Wang. 2017. A Century of\nScience: Globalization of Scientific Collaborations, Citations, and Innovations. In\nKDD \u201917 . ACM, 1437\u20131446.\n[5]Matthias Fey and Jan Eric Lenssen. 2019. Fast Graph Representation Learning\nwith PyTorch Geometric. ICLR 2019 Workshop: Representation Learning on Graphs\nand Manifolds (2019).\n[6]Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E.\nDahl. 2017. Neural Message Passing for Quantum Chemistry. In ICML . 1263\u20131272.\n[7]William L. Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive Represen-\ntation Learning on Large Graphs. In NeurIPS\u201917 .\n[8]Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep Residual\nLearning for Image Recognition. In CVPR\u201916 .\n[9]Thomas N. Kipf and Max Welling. 2017. Semi-Supervised Classification with\nGraph Convolutional Networks. In ICLR\u201917 .\n[10] Hang Li.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "96dc6fbf-2f14-4e07-ac4c-0f3385813d2b": {"__data__": {"id_": "96dc6fbf-2f14-4e07-ac4c-0f3385813d2b", "embedding": null, "metadata": {"page_label": "11", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the difference between inductive and transductive representation learning on large graphs?\n2. How does the use of residual learning improve image recognition performance?\n3. What is the role of self-attention in natural language processing?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer paper presented at the WWW '20 conference, which focuses on developing a multi-modal graph neural network for corporate relative valuation. The paper also mentions the FastGCN algorithm, which improves the efficiency of learning with Graph Convolutional Networks, and the Stochastic Training of Graph Convolu-tional Networks with Variance Reduction technique, which is significant in the field of graph neural networks. Additionally, the section references several other papers related to graph neural networks and representation learning.", "section_summary": "The section discusses various topics related to graph neural networks, including inductive and transductive representation learning on large graphs, the role of residual learning in image recognition, and the use of self-attention in natural language processing. The authors of the section include Hamilton, Ying, and Leskovec, He, Zhang, Ren, and Sun, Kipf and Welling, Li, Liu, Loshchilov and Hutter, Schlichtkrull et al., Shaw et al., Sinha et al., and Sun and Han. The section also mentions several papers and books, including \"Inductive Representation Learning on Large Graphs,\" \"Deep Residual Learning for Image Recognition,\" \"Semi-Supervised Classification with Graph Convolutional Networks,\" \"Learning to Rank for Information Retrieval and Natural Language Processing,\" \"SGDR: Stochastic Gradient Descent with Warm Restarts,\" \"Decoupled Weight Decay Regularization,\" \"Modeling Relational Data with Graph Convolutional Networks,\" \"Self-Attention with Relative Position Representations,\" \"An Overview of Microsoft Academic Service (MAS) and Ap", "excerpt_keywords": "Graph convolutional networks, semi-supervised classification, deep learning, neural networks, natural language processing, information retrieval, ranking, learning, optimization, regularization."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "f3ccaf27-46f6-4aaa-b10d-195f2156f4b6", "node_type": "4", "metadata": {"page_label": "11", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a8b06d28c38dad7b3eeb48534f0f6dd917aa3349a031c7bf425429fd635407d3"}, "2": {"node_id": "8a448c50-08b1-47f8-853b-919396983891", "node_type": "1", "metadata": {"page_label": "11", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "4193f4744201a0499d6bed904b345b1213475f8f2acab1f748ccc55fa4fa355d"}, "3": {"node_id": "15062f1b-c7f6-45d4-8ae4-7fedc8eda488", "node_type": "1", "metadata": {"page_label": "11", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "3e8fc50dc15c8f73919dc1191fb05765ef644b42ba6836c37457bd9f639a14be"}}, "hash": "7409ba8a8ce9e2c430770b45f2be1640f1e68830657d69f8bbd4a9ab86fb72c1", "text": "Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive Represen-\ntation Learning on Large Graphs. In NeurIPS\u201917 .\n[8]Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep Residual\nLearning for Image Recognition. In CVPR\u201916 .\n[9]Thomas N. Kipf and Max Welling. 2017. Semi-Supervised Classification with\nGraph Convolutional Networks. In ICLR\u201917 .\n[10] Hang Li. 2014. Learning to Rank for Information Retrieval and Natural Language\nProcessing, Second Edition . Morgan & Claypool Publishers. https://doi.org/10.\n2200/S00607ED2V01Y201410HLT026\n[11] Tie-Yan Liu. 2011. Learning to Rank for Information Retrieval . Springer.\n[12] Ilya Loshchilov and Frank Hutter. 2017. SGDR: Stochastic Gradient Descent with\nWarm Restarts. In ICLR\u201917 .\n[13] Ilya Loshchilov and Frank Hutter. 2019. Decoupled Weight Decay Regularization.\nInICLR\u201919 .\n[14] Michael Sejr Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg,\nIvan Titov, and Max Welling. 2018. Modeling Relational Data with Graph Convo-\nlutional Networks. In ESWC\u20192018 .\n[15] Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. 2018. Self-Attention with\nRelative Position Representations. In NAACL-HLT . 464\u2013468.\n[16] Arnab Sinha, Zhihong Shen, Yang Song, Hao Ma, Darrin Eide, Bo-June Paul Hsu,\nand Kuansan Wang. 2015. An Overview of Microsoft Academic Service (MAS)and Applications. In WWW Companion 2015 .\n[17] Yizhou Sun and Jiawei Han. 2012. Mining Heterogeneous Information", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "15062f1b-c7f6-45d4-8ae4-7fedc8eda488": {"__data__": {"id_": "15062f1b-c7f6-45d4-8ae4-7fedc8eda488", "embedding": null, "metadata": {"page_label": "11", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the difference between self-attention and graph attention networks?\n2. How can heterogeneous information networks be used for academic social network extraction and mining?\n3. What is the role of meta-path selection in clustering objects in heterogeneous information networks?", "prev_section_summary": "The section discusses various topics related to graph neural networks, including inductive and transductive representation learning on large graphs, the role of residual learning in image recognition, and the use of self-attention in natural language processing. The authors of the section include Hamilton, Ying, and Leskovec, He, Zhang, Ren, and Sun, Kipf and Welling, Li, Liu, Loshchilov and Hutter, Schlichtkrull et al., Shaw et al., Sinha et al., and Sun and Han. The section also mentions several papers and books, including \"Inductive Representation Learning on Large Graphs,\" \"Deep Residual Learning for Image Recognition,\" \"Semi-Supervised Classification with Graph Convolutional Networks,\" \"Learning to Rank for Information Retrieval and Natural Language Processing,\" \"SGDR: Stochastic Gradient Descent with Warm Restarts,\" \"Decoupled Weight Decay Regularization,\" \"Modeling Relational Data with Graph Convolutional Networks,\" \"Self-Attention with Relative Position Representations,\" \"An Overview of Microsoft Academic Service (MAS) and Ap", "section_summary": "The section discusses the use of heterogeneous information networks for academic social network extraction and mining. It also covers the role of meta-path selection in clustering objects in heterogeneous information networks. The section mentions several papers and research works related to this topic, including Self-Attention with Relative Position Representations, Microsoft Academic Service, Mining Heterogeneous Information Networks, Pathsim, and Arnet-miner. The section also discusses the difference between self-attention and graph attention networks, and provides an overview of the Attention is All You Need paper.", "excerpt_keywords": "1. Self-Attention,\n2. Relative Position Representations,\n3. Heterogeneous Information Networks,\n4. Pathsim,\n5. Meta-path selection,\n6. User-guided object clustering,\n7. Academic social networks,\n8. Graph Attention Networks,\n9. Heterogeneous Graph Attention,\n10. Attention is All you Need."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "f3ccaf27-46f6-4aaa-b10d-195f2156f4b6", "node_type": "4", "metadata": {"page_label": "11", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a8b06d28c38dad7b3eeb48534f0f6dd917aa3349a031c7bf425429fd635407d3"}, "2": {"node_id": "96dc6fbf-2f14-4e07-ac4c-0f3385813d2b", "node_type": "1", "metadata": {"page_label": "11", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "7409ba8a8ce9e2c430770b45f2be1640f1e68830657d69f8bbd4a9ab86fb72c1"}, "3": {"node_id": "4c5cb8e7-665e-410f-920d-33413ad7c183", "node_type": "1", "metadata": {"page_label": "11", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "2d6d1174409c645035140350372181fef32e1fc2e5b6e3e6a4f946b05f6057c9"}}, "hash": "3e8fc50dc15c8f73919dc1191fb05765ef644b42ba6836c37457bd9f639a14be", "text": "Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. 2018. Self-Attention with\nRelative Position Representations. In NAACL-HLT . 464\u2013468.\n[16] Arnab Sinha, Zhihong Shen, Yang Song, Hao Ma, Darrin Eide, Bo-June Paul Hsu,\nand Kuansan Wang. 2015. An Overview of Microsoft Academic Service (MAS)and Applications. In WWW Companion 2015 .\n[17] Yizhou Sun and Jiawei Han. 2012. Mining Heterogeneous Information Networks:\nPrinciples and Methodologies . Morgan & Claypool Publishers.\n[18] Yizhou Sun, Jiawei Han, Xifeng Yan, Philip S. Yu, and Tianyi Wu. 2011. Pathsim:\nMeta path-based top-k similarity search in heterogeneous information networks.\nInVLDB \u201911 .\n[19] Yizhou Sun, Brandon Norick, Jiawei Han, Xifeng Yan, Philip S. Yu, and Xiao\nYu. 2012. Integrating meta-path selection with user-guided object clustering in\nheterogeneous information networks. In KDD\u201912 .\n[20] Jie Tang, Jing Zhang, Limin Yao, Juanzi Li, Li Zhang, and Zhong Su. 2008. Arnet-\nminer: extraction and mining of academic social networks. In KDD .\n[21] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones,\nAidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is All\nyou Need. In NeurIPS\u201917 .\n[22] Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro\nLi\u00f2, and Yoshua Bengio. 2018. Graph Attention Networks. In ICLR\u201918 .\n[23] Xiao Wang, Houye Ji, Chuan Shi, Bai Wang, Yanfang Ye, Peng Cui, and Philip S.\nYu. 2019. Heterogeneous Graph Attention", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "4c5cb8e7-665e-410f-920d-33413ad7c183": {"__data__": {"id_": "4c5cb8e7-665e-410f-920d-33413ad7c183", "embedding": null, "metadata": {"page_label": "11", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the history of graph neural networks and their development in the field of natural language processing?\n2. How do heterogeneous graph attention networks differ from traditional graph neural networks and what are their advantages?\n3. What are some recent advancements in the field of graph neural networks and their applications in various domains?", "prev_section_summary": "The section discusses the use of heterogeneous information networks for academic social network extraction and mining. It also covers the role of meta-path selection in clustering objects in heterogeneous information networks. The section mentions several papers and research works related to this topic, including Self-Attention with Relative Position Representations, Microsoft Academic Service, Mining Heterogeneous Information Networks, Pathsim, and Arnet-miner. The section also discusses the difference between self-attention and graph attention networks, and provides an overview of the Attention is All You Need paper.", "section_summary": "The section discusses the history and development of graph neural networks in natural language processing, the differences between traditional and heterogeneous graph attention networks, recent advancements in the field, and their applications in various domains. Key topics include attention is all you need, graph attention networks, heterogeneous graph attention networks, transformers, XLNet, graph transformer networks, and heterogeneous graph neural networks. Entities mentioned include Lukasz Kaiser, Illia Polosukhin, Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Li\u00f2, Yoshua Bengio, Xiao Wang, Houye Ji, Chuan Shi, Bai Wang, Yanfang Ye, Peng Cui, Philip S. Yu, Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R\u00c3\u00b9mi Louf, Morgan Funtowicz, Jamie Brew, Zhilin Yang, Zihang Dai, Yiming Yang, Jaime G. Carbonell, Ruslan Salakhutdi-nov,", "excerpt_keywords": "attention, neural networks, graph attention networks, heterogeneous graph attention networks, transformers, natural language processing, language understanding, generalized autoregressive pretraining, graph transformer networks, heterogeneous graph neural networks."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "f3ccaf27-46f6-4aaa-b10d-195f2156f4b6", "node_type": "4", "metadata": {"page_label": "11", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a8b06d28c38dad7b3eeb48534f0f6dd917aa3349a031c7bf425429fd635407d3"}, "2": {"node_id": "15062f1b-c7f6-45d4-8ae4-7fedc8eda488", "node_type": "1", "metadata": {"page_label": "11", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "3e8fc50dc15c8f73919dc1191fb05765ef644b42ba6836c37457bd9f639a14be"}, "3": {"node_id": "f9034dee-5640-4342-9a2d-c72ad0780d4c", "node_type": "1", "metadata": {"page_label": "11", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "0c38ace2cff0650460187b948e01e9ed8076f663daa1a15228329535283081a0"}}, "hash": "2d6d1174409c645035140350372181fef32e1fc2e5b6e3e6a4f946b05f6057c9", "text": "Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is All\nyou Need. In NeurIPS\u201917 .\n[22] Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro\nLi\u00f2, and Yoshua Bengio. 2018. Graph Attention Networks. In ICLR\u201918 .\n[23] Xiao Wang, Houye Ji, Chuan Shi, Bai Wang, Yanfang Ye, Peng Cui, and Philip S.\nYu. 2019. Heterogeneous Graph Attention Network. In KDD\u201919 . 2022\u20132032.\n[24] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue,\nAnthony Moi, Pierric Cistac, Tim Rault, R\u00c3\u013emi Louf, Morgan Funtowicz, and\nJamie Brew. 2019. Transformers: State-of-the-art Natural Language Processing.\narXiv:cs.CL/1910.03771\n[25] Zhilin Yang, Zihang Dai, Yiming Yang, Jaime G. Carbonell, Ruslan Salakhutdi-\nnov, and Quoc V. Le. 2019. XLNet: Generalized Autoregressive Pretraining for\nLanguage Understanding. In NeurIPS\u201919 .\n[26] Seongjun Yun, Minbyul Jeong, Raehyun Kim, Jaewoo Kang, and Hyunwoo J. Kim.\n2019. Graph Transformer Networks. In NeurIPS\u201919 .\n[27] Chuxu Zhang, Dongjin Song, Chao Huang, Ananthram Swami, and Nitesh V.\nChawla. 2019. Heterogeneous Graph Neural Network. In WWW\u201919 .\n[28] Fanjin Zhang, Xiao Liu, Jie Tang, Yuxiao Dong, Peiran Yao, Jie Zhang, Xiaotao\nGu, Yan Wang, Bin Shao, Rui Li, and Kuansan Wang. 2019. OAG: Toward Linking\nLarge-scale Heterogeneous Entity Graphs. In KDD\u201919", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "f9034dee-5640-4342-9a2d-c72ad0780d4c": {"__data__": {"id_": "f9034dee-5640-4342-9a2d-c72ad0780d4c", "embedding": null, "metadata": {"page_label": "11", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the difference between Heterogeneous Graph Neural Network and OAG in linking large-scale heterogeneous entity graphs?\n2. How does Layer-Dependent Importance Sampling improve the training of deep and large graph convolutional networks?\n3. What is the significance of Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation in the field of graph neural networks?", "prev_section_summary": "The section discusses the history and development of graph neural networks in natural language processing, the differences between traditional and heterogeneous graph attention networks, recent advancements in the field, and their applications in various domains. Key topics include attention is all you need, graph attention networks, heterogeneous graph attention networks, transformers, XLNet, graph transformer networks, and heterogeneous graph neural networks. Entities mentioned include Lukasz Kaiser, Illia Polosukhin, Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Li\u00f2, Yoshua Bengio, Xiao Wang, Houye Ji, Chuan Shi, Bai Wang, Yanfang Ye, Peng Cui, Philip S. Yu, Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R\u00c3\u00b9mi Louf, Morgan Funtowicz, Jamie Brew, Zhilin Yang, Zihang Dai, Yiming Yang, Jaime G. Carbonell, Ruslan Salakhutdi-nov,", "section_summary": "The section discusses the difference between Heterogeneous Graph Neural Network (HGNN) and OAG in linking large-scale heterogeneous entity graphs. It also explains how Layer-Dependent Importance Sampling improves the training of deep and large graph convolutional networks. The significance of Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation in the field of graph neural networks is also highlighted. The authors of the section are Chuxu Zhang, Dongjin Song, Chao Huang, Ananthram Swami, and Nitesh V. Chawla, Fanjin Zhang, Xiao Liu, Jie Tang, Yuxiao Dong, Peiran Yao, Jie Zhang, Xiaotao Gu, Yan Wang, Bin Shao, Rui Li, and Kuansan Wang. The document was created on 2023-11-06, last modified on 2023-10-20, and last accessed on 2023-11-30. The title of the document is \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative", "excerpt_keywords": "1. Heterogeneous Graph Neural Network\n2. WWW'19\n3. Dongjin Song\n4. Chao Huang\n5. Ananthram Swami\n6. Nitesh V. Chawla\n7. OAG\n8. Linking\n9. Large-scale Heterogeneous Entity Graphs\n10. KDD'19"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "f3ccaf27-46f6-4aaa-b10d-195f2156f4b6", "node_type": "4", "metadata": {"page_label": "11", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a8b06d28c38dad7b3eeb48534f0f6dd917aa3349a031c7bf425429fd635407d3"}, "2": {"node_id": "4c5cb8e7-665e-410f-920d-33413ad7c183", "node_type": "1", "metadata": {"page_label": "11", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "2d6d1174409c645035140350372181fef32e1fc2e5b6e3e6a4f946b05f6057c9"}}, "hash": "0c38ace2cff0650460187b948e01e9ed8076f663daa1a15228329535283081a0", "text": ".\n[27] Chuxu Zhang, Dongjin Song, Chao Huang, Ananthram Swami, and Nitesh V.\nChawla. 2019. Heterogeneous Graph Neural Network. In WWW\u201919 .\n[28] Fanjin Zhang, Xiao Liu, Jie Tang, Yuxiao Dong, Peiran Yao, Jie Zhang, Xiaotao\nGu, Yan Wang, Bin Shao, Rui Li, and Kuansan Wang. 2019. OAG: Toward Linking\nLarge-scale Heterogeneous Entity Graphs. In KDD\u201919 .\n[29] Difan Zou, Ziniu Hu, Yewen Wang, Song Jiang, Yizhou Sun, and Quanquan Gu.\n2019. Layer-Dependent Importance Sampling for Training Deep and Large Graph\nConvolutional Networks. In NeurIPS\u201919 .", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "6c789761-7721-438b-83f6-32580146d47a": {"__data__": {"id_": "6c789761-7721-438b-83f6-32580146d47a", "embedding": null, "metadata": {"page_label": "93", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the relationship between network-oriented modeling and the dynamical systems perspective on mental processes?\n2. Can any adaptive dynamical system be modeled by a self-modeling network?\n3. What are the applications of a network-oriented modeling approach in scientific domains?", "prev_section_summary": "The section discusses the difference between Heterogeneous Graph Neural Network (HGNN) and OAG in linking large-scale heterogeneous entity graphs. It also explains how Layer-Dependent Importance Sampling improves the training of deep and large graph convolutional networks. The significance of Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation in the field of graph neural networks is also highlighted. The authors of the section are Chuxu Zhang, Dongjin Song, Chao Huang, Ananthram Swami, and Nitesh V. Chawla, Fanjin Zhang, Xiao Liu, Jie Tang, Yuxiao Dong, Peiran Yao, Jie Zhang, Xiaotao Gu, Yan Wang, Bin Shao, Rui Li, and Kuansan Wang. The document was created on 2023-11-06, last modified on 2023-10-20, and last accessed on 2023-11-30. The title of the document is \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative", "section_summary": "The section discusses the relationship between network-oriented modeling and the dynamical systems perspective on mental processes. It is mathematically proven that any dynamical system can be modeled as a temporal-causal network model and that any adaptive dynamical system (of any order) can be modeled by a self-modeling network (of the same order). The scope of applicability of a network-oriented modeling approach covers mental processes described by networks, social (interaction) processes described by social network models, and more. The section also mentions the longstanding tradition of network-oriented modeling in AI.", "excerpt_keywords": "Dynamical system, Adaptive, Network model, Self-modeling network, Mathematical analysis, Temporal-causal network, Causal pathways, Causal explanation, AI, Scientific domains."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "fb80ecac-79cd-47f7-a382-727a1c8a533e", "node_type": "4", "metadata": {"page_label": "93", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "b51f6693e2d1f769e18a27b56001b30d395b8eeaacf37fcff15cdcc858244c91"}, "3": {"node_id": "5243f89c-b9ab-442b-b607-6b8ea6277c0f", "node_type": "1", "metadata": {"page_label": "93", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "41bbd130de97741fe94b0536e2c8bc13fb9e6a9a19e78e92377803b76c8a53c3"}}, "hash": "156d3c031ec6488d878189fec780c8c6c7f0855d247fac0397ca2e0cc8b0e2ff", "text": "Cognitive Systems Research 70 (2021) 93\u2013100\nAvailable online 20 July 2021\n1389-0417/\u00a9 2021 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license ( http://creativecommons.org/licenses/by/4.0/ ).On the dynamics and adaptivity of mental processes: Relating adaptive \ndynamical systems and self-modeling network models by \nmathematical analysis \nJan Treur \nSocial AI Group, Department of Computer Science, Vrije Universiteit, Amsterdam, the Netherlands   \nARTICLE INFO  \nKeywords: \nDynamical system \nAdaptive \nNetwork model \nSelf-modeling network ABSTRACT  \nIn this paper, it is addressed by mathematical analysis how network-oriented modeling relates to the dynamical \nsystems perspective on mental processes. It has been mathematically proven that any dynamical system can be \nmodeled as a temporal-causal network model and that any adaptive dynamical system (of any order) can be \nmodeled by a self-modeling network (of the same order).   \n1.Introduction \nThe scope of applicability of a network-oriented modelling approach \ncovers mental processes described by networks for the interplay of \nmental states, social (interaction) processes described by social network \nmodels, and more; e.g., (Treur, 2016a; Treur, 2016b; Treur, 2017 ). In \nfact, any scientific area in which networks of causal relations and causal \npathways within them are used to describe theories, hypotheses and \nfindings falls within the scope of applicability of such a network- \noriented modeling approach. This covers practically all scientific do-\nmains, as causal explanation is used as a main vehicle almost every -\nwhere in science. Besides, as a modeling paradigm it has a longstanding \ntradition in AI; e.g.,", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "5243f89c-b9ab-442b-b607-6b8ea6277c0f": {"__data__": {"id_": "5243f89c-b9ab-442b-b607-6b8ea6277c0f", "embedding": null, "metadata": {"page_label": "93", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the role of causality in mental processes and how is it emphasized in philosophy of mind?\n2. How does network-oriented modeling approach apply to various scientific domains and what is its longstanding tradition in AI?\n3. What are the challenges in modeling realistic mental processes using a dynamical system perspective and how can network-oriented modeling address them?", "prev_section_summary": "The section discusses the relationship between network-oriented modeling and the dynamical systems perspective on mental processes. It is mathematically proven that any dynamical system can be modeled as a temporal-causal network model and that any adaptive dynamical system (of any order) can be modeled by a self-modeling network (of the same order). The scope of applicability of a network-oriented modeling approach covers mental processes described by networks, social (interaction) processes described by social network models, and more. The section also mentions the longstanding tradition of network-oriented modeling in AI.", "section_summary": "The section discusses the role of causality in mental processes and how it is emphasized in philosophy of mind. It also explores the application of network-oriented modeling to various scientific domains and its longstanding tradition in AI. The challenges in modeling realistic mental processes using a dynamical system perspective are addressed, and the importance of adaptivity in mental processes is highlighted. The section also discusses the limitations of network-oriented modeling in covering dynamics well and the need for a separate treatment of adaptivity.", "excerpt_keywords": "causality, network-oriented modeling, scientific domains, AI, philosophy of mind, mental processes, dynamical systems, state-determined systems, Ashby, Beer, Kelso, Scherer, Thelen, Smith, adaptivity, temporal dynamics, network nodes, states, applicability."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "fb80ecac-79cd-47f7-a382-727a1c8a533e", "node_type": "4", "metadata": {"page_label": "93", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "b51f6693e2d1f769e18a27b56001b30d395b8eeaacf37fcff15cdcc858244c91"}, "2": {"node_id": "6c789761-7721-438b-83f6-32580146d47a", "node_type": "1", "metadata": {"page_label": "93", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "156d3c031ec6488d878189fec780c8c6c7f0855d247fac0397ca2e0cc8b0e2ff"}, "3": {"node_id": "ee5db39c-b21e-4726-9736-122a98b25b66", "node_type": "1", "metadata": {"page_label": "93", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "e49490c6bcd37155ac820840348febe5bf31f68b9921276910b99d9ad1bc741e"}}, "hash": "41bbd130de97741fe94b0536e2c8bc13fb9e6a9a19e78e92377803b76c8a53c3", "text": "and more; e.g., (Treur, 2016a; Treur, 2016b; Treur, 2017 ). In \nfact, any scientific area in which networks of causal relations and causal \npathways within them are used to describe theories, hypotheses and \nfindings falls within the scope of applicability of such a network- \noriented modeling approach. This covers practically all scientific do-\nmains, as causal explanation is used as a main vehicle almost every -\nwhere in science. Besides, as a modeling paradigm it has a longstanding \ntradition in AI; e.g., (Kuipers, 1984; Kuipers and Kassirer, 1983; Pearl, \n2000 ). Also, from the area of philosophy of mind the important role of \ncausality in mental processes is emphasized; e.g., (Kim, 1996 ). \nIn (Port and Van Gelder, 1995; van Gelder, 1998; van Gelder and \nPort, 1995 ) it is argued that modeling realistic mental processes needs a \ndynamical system perspective where the temporal dimension is covered \nwell; here a dynamical system is defined by the notion of a state- \ndetermined system as analysed by (Ashby, 1960 ). For similar positions \nemphasizing the role of dynamics to model mental processes, see, for \nexample (Beer, 2000; Kelso, 1995; Scherer, 2009; Thelen and Smith, \n1994 ). Given these positions, an important question becomes in how far \nnetwork-oriented modeling is able to cover dynamics well. Moreover, \nalso adaptivity of mental processes is an important issue in this, what \nsometimes is silently assumed to be covered by dynamics as well, but \ndeserves its own treatment. In the first place, the way in which the temporal dynamics of the \nimpacts of network nodes (also called states) on other nodes are handled \nin (Treur, 2016a; Treur, 2016b ) is important here, as for many appli -\ncations detailed handling of", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "ee5db39c-b21e-4726-9736-122a98b25b66": {"__data__": {"id_": "ee5db39c-b21e-4726-9736-122a98b25b66", "embedding": null, "metadata": {"page_label": "93", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. How well can network-oriented modeling cover the dynamics and adaptivity of mental processes?\n2. What are the mathematical theorems presented in the current paper that provide a positive answer to this question?\n3. Can any adaptive dynamical system be modeled by a self-modeling network model?", "prev_section_summary": "The section discusses the role of causality in mental processes and how it is emphasized in philosophy of mind. It also explores the application of network-oriented modeling to various scientific domains and its longstanding tradition in AI. The challenges in modeling realistic mental processes using a dynamical system perspective are addressed, and the importance of adaptivity in mental processes is highlighted. The section also discusses the limitations of network-oriented modeling in covering dynamics well and the need for a separate treatment of adaptivity.", "section_summary": "The section discusses the use of network-oriented modeling in understanding the dynamics and adaptivity of mental processes. The author highlights the importance of handling temporal dynamics of network nodes and the adaptivity of network models. The section also mentions two theorems presented in the current paper that provide a positive answer to the question of how well dynamics is covered by network-oriented modeling. The first theorem shows that temporal-causal network models can model any dynamical system, while the second theorem shows that any adaptive dynamical system can be modeled by a self-modeling network model.", "excerpt_keywords": "1. Network-oriented modeling\n2. Dynamics\n3. Adaptivity\n4. Temporal causality\n5. First-order equations\n6. Adaptive differential equations\n7. Self-modeling network models\n8. Mathematical analysis\n9. Positive answer\n10. Temporal dynamics of impacts"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "fb80ecac-79cd-47f7-a382-727a1c8a533e", "node_type": "4", "metadata": {"page_label": "93", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "b51f6693e2d1f769e18a27b56001b30d395b8eeaacf37fcff15cdcc858244c91"}, "2": {"node_id": "5243f89c-b9ab-442b-b607-6b8ea6277c0f", "node_type": "1", "metadata": {"page_label": "93", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "41bbd130de97741fe94b0536e2c8bc13fb9e6a9a19e78e92377803b76c8a53c3"}, "3": {"node_id": "2b8a9431-524f-4944-8ab6-cc5042aed4e8", "node_type": "1", "metadata": {"page_label": "93", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "5b6b07c565d0f75b40d29815bc99c83ed509e5165065a0f956de4860f71b66c5"}}, "hash": "e49490c6bcd37155ac820840348febe5bf31f68b9921276910b99d9ad1bc741e", "text": "Given these positions, an important question becomes in how far \nnetwork-oriented modeling is able to cover dynamics well. Moreover, \nalso adaptivity of mental processes is an important issue in this, what \nsometimes is silently assumed to be covered by dynamics as well, but \ndeserves its own treatment. In the first place, the way in which the temporal dynamics of the \nimpacts of network nodes (also called states) on other nodes are handled \nin (Treur, 2016a; Treur, 2016b ) is important here, as for many appli -\ncations detailed handling of such dynamics is required, for example, \nsince usually there is no global synchrony of everything that happens in \nthe world or in the mind. Secondly, in addition to such dynamics of \nimpacts, in (Treur, 2020 ) adaptivity of network models is addressed as \nwell. This provides good application possibilities for the many applica -\ntions that concern learning or other forms of adaptation. However, these \nare just indications or expectations and some experiences; in addition to \nthis, the question on how well dynamics is covered by network-oriented \nmodeling can also be analysed mathematically. This is what is done in \nthe current paper. Two theorems are presented that provide a positive \nanswer to this question, one for nonadaptive dynamical systems in \nrelation to network models and one for adaptive dynamical systems in \nrelation to self-modeling network models. \nMore specifically, in this paper, in Section 5 it is shown mathemat -\nically that temporal-causal network models can model any dynamical \nsystem (Ashby, 1960 ), or, equivalently, any set of first-order equations. \nNext, in Section 6 it is shown how this also holds for any adaptive \ndynamical system or any adaptive set of first-order differential equa-\ntions. It is shown in particular how any adaptive dynamical system can \nbe modeled by a self-modeling network model. But", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "2b8a9431-524f-4944-8ab6-cc5042aed4e8": {"__data__": {"id_": "2b8a9431-524f-4944-8ab6-cc5042aed4e8", "embedding": null, "metadata": {"page_label": "93", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the relationship between temporal-causal network models and adaptive dynamical systems in modeling any set of first-order equations?\n2. How does the notion of state-determined system from Ashby (1960) and Port and Van Gelder (1995) relate to the notion of dynamical system used in this paper?\n3. What is the mathematical analysis presented in this paper that shows how any adaptive dynamical system can be modeled by a self-modeling network model?", "prev_section_summary": "The section discusses the use of network-oriented modeling in understanding the dynamics and adaptivity of mental processes. The author highlights the importance of handling temporal dynamics of network nodes and the adaptivity of network models. The section also mentions two theorems presented in the current paper that provide a positive answer to the question of how well dynamics is covered by network-oriented modeling. The first theorem shows that temporal-causal network models can model any dynamical system, while the second theorem shows that any adaptive dynamical system can be modeled by a self-modeling network model.", "section_summary": "The section discusses the relationship between temporal-causal network models and adaptive dynamical systems in modeling any set of first-order equations. The authors show mathematically that temporal-causal network models can model any dynamical system or any set of first-order differential equations. They also show how any adaptive dynamical system can be modeled by a self-modeling network model. The authors discuss the notion of state-determined system from Ashby (1960) and Port and Van Gelder (1995) and show that it is equivalent to a set of first-order differential equations. The authors also provide a mathematical analysis that shows how any adaptive dynamical system can be modeled by a self-modeling network model.", "excerpt_keywords": "1. Temporal-causal network models\n2. Dynamical systems\n3. First-order equations\n4. Adaptive dynamical systems\n5. Self-modeling network models\n6. State-determined systems\n7. First-order differential equations\n8. Equivalence\n9. Mathematical modeling\n10. Adaptive sets of differential equations"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "fb80ecac-79cd-47f7-a382-727a1c8a533e", "node_type": "4", "metadata": {"page_label": "93", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "b51f6693e2d1f769e18a27b56001b30d395b8eeaacf37fcff15cdcc858244c91"}, "2": {"node_id": "ee5db39c-b21e-4726-9736-122a98b25b66", "node_type": "1", "metadata": {"page_label": "93", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "e49490c6bcd37155ac820840348febe5bf31f68b9921276910b99d9ad1bc741e"}}, "hash": "5b6b07c565d0f75b40d29815bc99c83ed509e5165065a0f956de4860f71b66c5", "text": "relation to self-modeling network models. \nMore specifically, in this paper, in Section 5 it is shown mathemat -\nically that temporal-causal network models can model any dynamical \nsystem (Ashby, 1960 ), or, equivalently, any set of first-order equations. \nNext, in Section 6 it is shown how this also holds for any adaptive \ndynamical system or any adaptive set of first-order differential equa-\ntions. It is shown in particular how any adaptive dynamical system can \nbe modeled by a self-modeling network model. But first, some prepa -\nrations are made. In Section 2 the notion of state-determined system \nfrom (Ashby, 1960; Port and Van Gelder, 1995 ) is discussed which is the \nbasis for the notion of dynamical system used. In Section 3, it is shown \nthat this notion is equivalent to a set of first-order differential equations. \nE-mail address: j.treur@vu.nl .  \nContents lists available at ScienceDirect \nCognitive Systems Research \njournal homepag e: www.else vier.com/loc ate/cogs ys \nhttps://doi.org/10.1016/j.cogsys.2021.07.003 \nReceived 30 April 2021; Accepted 13 July 2021", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "a42737a5-ac14-4639-b0a9-050b2bc5b0c2": {"__data__": {"id_": "a42737a5-ac14-4639-b0a9-050b2bc5b0c2", "embedding": null, "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the definition of a state-determined system and how does it relate to the notion of a dynamical system?\n2. What are the three features of a state-determined system and how are they reflected in modeling and simulation?\n3. How does the state-determined system assumption relate to the notions of temporal factorisation and criterial causation?", "prev_section_summary": "The section discusses the relationship between temporal-causal network models and adaptive dynamical systems in modeling any set of first-order equations. The authors show mathematically that temporal-causal network models can model any dynamical system or any set of first-order differential equations. They also show how any adaptive dynamical system can be modeled by a self-modeling network model. The authors discuss the notion of state-determined system from Ashby (1960) and Port and Van Gelder (1995) and show that it is equivalent to a set of first-order differential equations. The authors also provide a mathematical analysis that shows how any adaptive dynamical system can be modeled by a self-modeling network model.", "section_summary": "The section discusses the concept of a state-determined system and its relation to the notion of a dynamical system. The state-determined system assumption is adopted from Ashby (1960) and taken as the basis to describe what a dynamical system is. Three features of a state-determined system are discussed: the future behavior cannot depend on states the system might have been in before the current state, the current state determines future behavior implies the existence of some rule of evolution describing the behavior of the system as a function of its current state, and future behaviors are uniquely determined. The section also relates the state-determined system assumption to the notions of temporal factorization and criterial causation.", "excerpt_keywords": "state-determined system, dynamical system, modeling, simulation, rule of evolution, differential equations, network models, temporal factorization, criterial causation, past history, current state, future behavior, uniqueness, state space sequences, fork, prediction, simulation, rule of evolution, specification, succinct, useful, Ashby, Port, van Gelder, Treur."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "74571bde-6fcb-4af9-ab8a-35c53b325152", "node_type": "4", "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "d2baa62743b351ba26cd4e7fdac22c4c69777d50fd09182b343bd5c24b7c7286"}, "3": {"node_id": "b8e3088d-2326-4b7c-b6e7-6027432a2844", "node_type": "1", "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "498bb4fc7539838520c9a7c90a38a5dee2bfaf40da036313272a51177ebdd4d3"}}, "hash": "a85f3d193333fac3bbc5de3ef46322e1d860212db3481334dddfe079ef9d7ec0", "text": "Cognitive Systems Research 70 (2021) 93\u2013100\n94Section 4 briefly summarises the self-modeling network modeling \napproach. Finally, Section 7 is a discussion. \n2.The State-Determined system assumption \nThe notion of state-determined system , adopted from Ashby (1960) is \ntaken as the basis to describe what a dynamical system is in (van Gelder \nand Port, 1995 , p. 6). That a system is state-determined means that its \ncurrent state always determines a unique future behaviour . This property is \nreflected in modelling and simulation. Three features in particular are \n(van Gelder and Port, 1995 ):  \n\u2022The future behaviour cannot depend on states the system might have \nbeen in before the current state: past history only can make a dif-\nference insofar as it has left an effect on the current state. This means \nthat if you want to make a prediction on a next state, for example by \nsimulation, only the information from the current state is needed, not \nfrom earlier states. \n\u2022That the current state determines future behaviour implies the ex-\nistence of some rule of evolution describing the behaviour of the \nsystem as a function of its current state. The idea is that this rule can \nbe specified in some reasonable succinct and useful fashion. The \nformats of differential equations (Ashby, 1960; Port and van Gelder, \n1995 ) and of network models as described in (Treur, 2016a; Treur, \n2016b ) are examples of formats in which such rules of evolution can \nbe expressed.  \n\u2022That future behaviours are uniquely determined means that state \nspace sequences can never fork. This means that when a next state is \ndetermined out of a current state, there is only one outcome. \nFrom a wider philosophical perspective, this notion of state- \ndetermined system can be related to the notions temporal factorisation \nand criterial causation as discussed in", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "b8e3088d-2326-4b7c-b6e7-6027432a2844": {"__data__": {"id_": "b8e3088d-2326-4b7c-b6e7-6027432a2844", "embedding": null, "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the relationship between adaptive dynamical systems and self-modeling network models in the context of mental processes?\n2. How do mathematical analysis techniques help to understand the dynamics and adaptivity of mental processes?\n3. What is the role of state-determined systems in the study of mental processes, and how does the choice of state properties affect their practical usability?", "prev_section_summary": "The section discusses the concept of a state-determined system and its relation to the notion of a dynamical system. The state-determined system assumption is adopted from Ashby (1960) and taken as the basis to describe what a dynamical system is. Three features of a state-determined system are discussed: the future behavior cannot depend on states the system might have been in before the current state, the current state determines future behavior implies the existence of some rule of evolution describing the behavior of the system as a function of its current state, and future behaviors are uniquely determined. The section also relates the state-determined system assumption to the notions of temporal factorization and criterial causation.", "section_summary": "The section discusses the relationship between adaptive dynamical systems and self-modeling network models in the context of mental processes. Mathematical analysis techniques are used to understand the dynamics and adaptivity of mental processes. State-determined systems play a crucial role in the study of mental processes, and the choice of state properties affects their practical usability. The section also discusses the notion of temporal factorisation and criterial causation, and the possibility of a choice of a proper set of state properties to obtain a state-determined system that is practically usable.", "excerpt_keywords": "1. Dynamical Systems Theory\n2. State-determined system\n3. Temporal factorisation\n4. Criterial causation\n5. Choice of state properties\n6. Practical usability\n7. Newton's gravitation laws\n8. Locality assumption\n9. Approximation\n10. Useful outcome"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "74571bde-6fcb-4af9-ab8a-35c53b325152", "node_type": "4", "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "d2baa62743b351ba26cd4e7fdac22c4c69777d50fd09182b343bd5c24b7c7286"}, "2": {"node_id": "a42737a5-ac14-4639-b0a9-050b2bc5b0c2", "node_type": "1", "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a85f3d193333fac3bbc5de3ef46322e1d860212db3481334dddfe079ef9d7ec0"}, "3": {"node_id": "b7294d1e-6da7-4f44-ad24-01f6c545c93f", "node_type": "1", "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "5521ad6ac58dbb10630a6a42a3a0bc325642f115c976b9ad06a880ac9733dcf5"}}, "hash": "498bb4fc7539838520c9a7c90a38a5dee2bfaf40da036313272a51177ebdd4d3", "text": "and van Gelder, \n1995 ) and of network models as described in (Treur, 2016a; Treur, \n2016b ) are examples of formats in which such rules of evolution can \nbe expressed.  \n\u2022That future behaviours are uniquely determined means that state \nspace sequences can never fork. This means that when a next state is \ndetermined out of a current state, there is only one outcome. \nFrom a wider philosophical perspective, this notion of state- \ndetermined system can be related to the notions temporal factorisation \nand criterial causation as discussed in (Treur, 2007a; Treur, 2007b; \nTreur, 2021; Tse, 2013 ). \nThe possibility of a choice of a proper set of state properties is the \ncrucial factor to obtain a state-determined system that is practically \nusable. The validity of the assumptions underlying the Dynamical Sys-\ntems Theory depends on the existence of such sets. For example, if to \nobtain a proper state-determined system to study some mental process, \nall states of the universe (including, for example the positions of all \nplanets and stars and even the mental states of all other humans) are \nneeded, then for practical purposes this perspective is useless. The truth \nis that, even for those who believe in science, for example, concerning \nNewton \u2019s gravitation laws, even for the application of such solid \nmonumental laws, in principle all mass and positions from the universe \nhave to be taken into account, which obviously is infeasible. But in \npractice, only mass that is not very far away is incorporated in a model, \nwhich then in fact provides an approximation. If such an approximation \nis accurate enough (objects very far away have some effect, but this is a \nvery small effect), then still a useful outcome can be obtained. So, more \nin general, usually an additional kind of locality assumption is made that \nto model", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "b7294d1e-6da7-4f44-ad24-01f6c545c93f": {"__data__": {"id_": "b7294d1e-6da7-4f44-ad24-01f6c545c93f", "embedding": null, "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the difference between adaptive dynamical systems and self-modeling network models in the context of mathematical analysis?\n2. How does the locality assumption play a role in modeling specific processes in science?\n3. What are some common formats used to describe dynamical systems adequately, and how do they differ from each other?", "prev_section_summary": "The section discusses the relationship between adaptive dynamical systems and self-modeling network models in the context of mental processes. Mathematical analysis techniques are used to understand the dynamics and adaptivity of mental processes. State-determined systems play a crucial role in the study of mental processes, and the choice of state properties affects their practical usability. The section also discusses the notion of temporal factorisation and criterial causation, and the possibility of a choice of a proper set of state properties to obtain a state-determined system that is practically usable.", "section_summary": "The section discusses the difference between adaptive dynamical systems and self-modeling network models in the context of mathematical analysis. It also explores the role of the locality assumption in modeling specific processes in science and the formats used to describe dynamical systems adequately. The section highlights the importance of finding a larger set of state variables that includes the given variables and is state-determined to obtain a useful outcome. The paper analyzes the formats in which such dynamical systems can be described adequately, including sets of first-order differential equations and the temporal-causal network format. The section also discusses adaptive dynamical systems and self-modeling networks.", "excerpt_keywords": "dynamical systems, state-determined, approximation, locality assumption, state variables, Ashby, hypothesis, first-order differential equations, temporal-causal network, adaptive dynamical systems, self-modeling networks."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "74571bde-6fcb-4af9-ab8a-35c53b325152", "node_type": "4", "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "d2baa62743b351ba26cd4e7fdac22c4c69777d50fd09182b343bd5c24b7c7286"}, "2": {"node_id": "b8e3088d-2326-4b7c-b6e7-6027432a2844", "node_type": "1", "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "498bb4fc7539838520c9a7c90a38a5dee2bfaf40da036313272a51177ebdd4d3"}, "3": {"node_id": "d5e7c57b-e43f-4036-a5db-b580d97735d2", "node_type": "1", "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "ae8b4959804b5fb46900dbf68d2c4cfa07c37e38333b58392c00e0bdddfc1a14"}}, "hash": "5521ad6ac58dbb10630a6a42a3a0bc325642f115c976b9ad06a880ac9733dcf5", "text": "even for the application of such solid \nmonumental laws, in principle all mass and positions from the universe \nhave to be taken into account, which obviously is infeasible. But in \npractice, only mass that is not very far away is incorporated in a model, \nwhich then in fact provides an approximation. If such an approximation \nis accurate enough (objects very far away have some effect, but this is a \nvery small effect), then still a useful outcome can be obtained. So, more \nin general, usually an additional kind of locality assumption is made that \nto model a specific process (and not the whole universe), a limited set of \nstate variables can be found to get a state-determined system. In Ashby \n(1960), such a hypothesis is expressed as follows: \n\u2018Because of its importance, science searches persistently for the state- \ndetermined. As a working guide, the scientist has for some centuries \nfollowed the hypothesis that, given a set of variables, he can always \nfind a larger set that (1) includes the given variables, and (2) is state- \ndetermined. Much research work consists of trying to identify such a \nlarger set, for when it is too small, important variables will be left out \nof the account, and the behaviour of the set will be capricious. The \nassumption that such a larger set exists is implicit in almost all sci-\nence, but, being fundamental, it is seldom mentioned explicitly. \u2019 \n(Ashby, 1960 , p. 28). In this paper it will be analyzed in some more depth in what formats \nin general such dynamical systems can be described adequately. It will \nturn out that one such format is by sets of first-order differential equa-\ntions, and another adequate format is the temporal-causal network \nformat described in (Treur, 2016a; Treur, 2016a ). Similarly, adaptive \ndynamical systems and self-modeling networks", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "d5e7c57b-e43f-4036-a5db-b580d97735d2": {"__data__": {"id_": "d5e7c57b-e43f-4036-a5db-b580d97735d2", "embedding": null, "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What are some common mathematical formats used to describe dynamical systems?\n2. How can adaptive processes be adequately described using mathematical formats?\n3. What are the criteria for a dynamical system and how can they be formalized in a numerical manner?", "prev_section_summary": "The section discusses the difference between adaptive dynamical systems and self-modeling network models in the context of mathematical analysis. It also explores the role of the locality assumption in modeling specific processes in science and the formats used to describe dynamical systems adequately. The section highlights the importance of finding a larger set of state variables that includes the given variables and is state-determined to obtain a useful outcome. The paper analyzes the formats in which such dynamical systems can be described adequately, including sets of first-order differential equations and the temporal-causal network format. The section also discusses adaptive dynamical systems and self-modeling networks.", "section_summary": "The section discusses the mathematical formats used to describe dynamical systems and adaptive processes. It highlights the adequacy of sets of first-order differential equations and the temporal-causal network format for describing dynamical systems. Additionally, adaptive dynamical systems and self-modeling networks are also mentioned as adequate formats for adaptive processes. The section also provides a detailed explanation of the criteria for a dynamical system and how they can be formalized in a numerical manner.", "excerpt_keywords": "dynamical systems, first-order differential equations, sets of equations, temporal-causal network, adaptive processes, self-modeling networks, mathematical formats, state variables, time t, rule of evolution, Fi function, Xi(t), Xn(t), s, continuous processes, smoothness, uniqueness."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "74571bde-6fcb-4af9-ab8a-35c53b325152", "node_type": "4", "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "d2baa62743b351ba26cd4e7fdac22c4c69777d50fd09182b343bd5c24b7c7286"}, "2": {"node_id": "b7294d1e-6da7-4f44-ad24-01f6c545c93f", "node_type": "1", "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "5521ad6ac58dbb10630a6a42a3a0bc325642f115c976b9ad06a880ac9733dcf5"}, "3": {"node_id": "fa64ff21-9041-4fa6-bf55-b326d25811e8", "node_type": "1", "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "c6774a8f38b5c6e1febf4370e2054e5884c1fccadf935ccbf51051af6535b186"}}, "hash": "ae8b4959804b5fb46900dbf68d2c4cfa07c37e38333b58392c00e0bdddfc1a14", "text": "in almost all sci-\nence, but, being fundamental, it is seldom mentioned explicitly. \u2019 \n(Ashby, 1960 , p. 28). In this paper it will be analyzed in some more depth in what formats \nin general such dynamical systems can be described adequately. It will \nturn out that one such format is by sets of first-order differential equa-\ntions, and another adequate format is the temporal-causal network \nformat described in (Treur, 2016a; Treur, 2016a ). Similarly, adaptive \ndynamical systems and self-modeling networks (Treur, 2020a; Treur, \n2020a ) turn out both to be adequate formats for adaptive processes. \n3.Dynamical systems and First-Order differential equations \nDynamical systems can be specified in mathematical formats; see \n(Ashby, 1960 , pp. 241\u2013252) for some details. In the first place a finite set \nof states (or state variables) X1, \u2026, Xn is assumed describing how the \nsystem changes over time via functions X1(t), \u2026, Xn(t) of time t. As \ndiscussed in Section 2, the criteria for a dynamical system can be \nformalized in a numerical manner by a relation (rule of evolution) that \nexpresses that for each time point t the future value of each state Xi at \ntime t +s uniquely depends on s and on X1(t), \u2026, Xn(t) and hence can be \ndescribed via some function Fi(X1, \u2026, Xn, s) in the following manner (see \nalso Ashby, 1960 , pp. 243\u2013244): \nXi(t+s)=Fi(X1(t),\u22ef,Xn(t),s) (1)  \nfors\u22650 \nAssuming continuous processes and smoothness (being differen -\ntiable) of the functions Xi(t) and Fi, these relations can", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "fa64ff21-9041-4fa6-bf55-b326d25811e8": {"__data__": {"id_": "fa64ff21-9041-4fa6-bf55-b326d25811e8", "embedding": null, "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the relationship between adaptive dynamical systems and self-modeling network models in the context of mental processes?\n2. How can the dynamics and adaptivity of mental processes be described using mathematical analysis and graph neural networks?\n3. What are the key components of a network model and how do they relate to the modeling approach described in Treur (2016a)?", "prev_section_summary": "The section discusses the mathematical formats used to describe dynamical systems and adaptive processes. It highlights the adequacy of sets of first-order differential equations and the temporal-causal network format for describing dynamical systems. Additionally, adaptive dynamical systems and self-modeling networks are also mentioned as adequate formats for adaptive processes. The section also provides a detailed explanation of the criteria for a dynamical system and how they can be formalized in a numerical manner.", "section_summary": "The section discusses the relationship between adaptive dynamical systems and self-modeling network models in the context of mental processes. It describes how the dynamics and adaptivity of mental processes can be described using mathematical analysis and graph neural networks. The section also outlines the key components of a network model and how they relate to the modeling approach described in Treur (2016a). The key topics and entities discussed in the section include adaptive dynamical systems, self-modeling network models, graph neural networks, first-order differential equations, state-determined systems, network models, connectivity characteristics, aggregation characteristics, timing characteristics, and speed factors.", "excerpt_keywords": "1. Network modeling\n2. State-determined systems\n3. First-order differential equations\n4. Continuous processes\n5. Smoothness\n6. Connectivity characteristics\n7. Aggregation characteristics\n8. Timing characteristics\n9. Self-modeling\n10. Network-oriented modeling"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "74571bde-6fcb-4af9-ab8a-35c53b325152", "node_type": "4", "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "d2baa62743b351ba26cd4e7fdac22c4c69777d50fd09182b343bd5c24b7c7286"}, "2": {"node_id": "d5e7c57b-e43f-4036-a5db-b580d97735d2", "node_type": "1", "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "ae8b4959804b5fb46900dbf68d2c4cfa07c37e38333b58392c00e0bdddfc1a14"}, "3": {"node_id": "29fc01a6-e599-43ea-bcc1-2787e7698c86", "node_type": "1", "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "4f5ad3bbbe332963d4a269072a738a8323c8f22f04e1c04538ef070c1d363f9c"}}, "hash": "c6774a8f38b5c6e1febf4370e2054e5884c1fccadf935ccbf51051af6535b186", "text": "uniquely depends on s and on X1(t), \u2026, Xn(t) and hence can be \ndescribed via some function Fi(X1, \u2026, Xn, s) in the following manner (see \nalso Ashby, 1960 , pp. 243\u2013244): \nXi(t+s)=Fi(X1(t),\u22ef,Xn(t),s) (1)  \nfors\u22650 \nAssuming continuous processes and smoothness (being differen -\ntiable) of the functions Xi(t) and Fi, these relations can be reformulated \n(see Box 1) into a set of first-order differential equations of the form \ndXi(t)\ndt=fi(X1(t),\u22ef,Xn(t)) (2) \nfor some functions fi(X1, \u2026, Xn); see also (Ashby, 1960 ), pp. 244\u2013246. \nNote that Xi may also occur in fi(X1, \u2026, Xn). Conversely, such a set of \nfirst-order differential equations always describes a state-determined \nsystem; so for the smooth continuous numerical case, state-determined \nsystems are the systems that can be described by sets of first-order dif-\nferential equations (Ashby, 1960 , p. 246). \n4.Self-Modeling network modeling \nAccording to the network-oriented modeling approach described in \n(Treur, 2016a ) a network model is characterised by:  \n\u2022connectivity characteristics \nConnections from a node (or state) X to a node Y and their weights \u03c9X,Y  \n\u2022aggregation characteristics \nFor any node Y, some combination function cY(..) defines aggregation \nthat is applied to the single impacts \u03c9X,YX(t) on Y through its \nincoming connections from states X  \n\u2022timing characteristics \nEach node Y has a speed factor \u03b7Y", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "29fc01a6-e599-43ea-bcc1-2787e7698c86": {"__data__": {"id_": "29fc01a6-e599-43ea-bcc1-2787e7698c86", "embedding": null, "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the mathematical formula for simulating the dynamics of a network model with connectivity, aggregation, and timing characteristics?\n2. What are the basic combination functions used in network-oriented modeling, and how do they enable the design of network models?\n3. How does the network model described in (Treur, 2016a) incorporate connectivity, aggregation, and timing characteristics to simulate and analyze network dynamics?", "prev_section_summary": "The section discusses the relationship between adaptive dynamical systems and self-modeling network models in the context of mental processes. It describes how the dynamics and adaptivity of mental processes can be described using mathematical analysis and graph neural networks. The section also outlines the key components of a network model and how they relate to the modeling approach described in Treur (2016a). The key topics and entities discussed in the section include adaptive dynamical systems, self-modeling network models, graph neural networks, first-order differential equations, state-determined systems, network models, connectivity characteristics, aggregation characteristics, timing characteristics, and speed factors.", "section_summary": "The section discusses the mathematical formula for simulating the dynamics of a network model with connectivity, aggregation, and timing characteristics. It also describes the basic combination functions used in network-oriented modeling and how they enable the design of network models. The network model described in (Treur, 2016a) incorporates connectivity, aggregation, and timing characteristics to simulate and analyze network dynamics. The difference (or differential) equations that are useful for simulation and analysis of network dynamics are also discussed.", "excerpt_keywords": "network model, connectivity characteristics, aggregation characteristics, timing characteristics, differential equations, simulation, analysis, network dynamics, declarative manner, mathematical functions, relations."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "74571bde-6fcb-4af9-ab8a-35c53b325152", "node_type": "4", "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "d2baa62743b351ba26cd4e7fdac22c4c69777d50fd09182b343bd5c24b7c7286"}, "2": {"node_id": "fa64ff21-9041-4fa6-bf55-b326d25811e8", "node_type": "1", "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "c6774a8f38b5c6e1febf4370e2054e5884c1fccadf935ccbf51051af6535b186"}, "3": {"node_id": "34f3377d-b7e2-4dd9-917b-b5cd9adcfcb7", "node_type": "1", "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "09157afd1a3f5fbf1b308ff680aea36d887805195e44f86b4efdcf29ff93d707"}}, "hash": "4f5ad3bbbe332963d4a269072a738a8323c8f22f04e1c04538ef070c1d363f9c", "text": "network-oriented modeling approach described in \n(Treur, 2016a ) a network model is characterised by:  \n\u2022connectivity characteristics \nConnections from a node (or state) X to a node Y and their weights \u03c9X,Y  \n\u2022aggregation characteristics \nFor any node Y, some combination function cY(..) defines aggregation \nthat is applied to the single impacts \u03c9X,YX(t) on Y through its \nincoming connections from states X  \n\u2022timing characteristics \nEach node Y has a speed factor \u03b7Y defining how fast it changes for \ngiven (aggregated) impact \nThe difference (or differential) equations that are useful for simula -\ntion purposes and also for analysis of network dynamics incorporate \nthese network characteristics \u03c9X,Y, cY(..), \u03b7Y: it holds \nY(t+\u0394t)=Y(t)+\u03b7Y[cY(\u03c9X1,YX1(t),\u22ef,\u03c9Xk,YXk(t))\u0000Y(t)]\u0394t  \ndY(t)\ndt=\u03b7Y[cY(\u03c9X1,YX1(t),\u22ef,\u03c9Xk,YXk(t))\u0000Y(t)] (3) \nfor any state Y and where X1,\u22ef,Xk are the states from which it gets \nits incoming connections. The above concepts enable to design network \nmodels and their dynamics in a declarative manner, based on mathe -\nmatically defined functions and relations. \nTo support the design of network models, for any application from a \nlibrary predefined basic combination functions bcfi(..), i =1, .., m are J. Treur", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "34f3377d-b7e2-4dd9-917b-b5cd9adcfcb7": {"__data__": {"id_": "34f3377d-b7e2-4dd9-917b-b5cd9adcfcb7", "embedding": null, "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. How do heterogeneous multi-modal graph neural networks enable the design of network models and their dynamics in a declarative manner?\n2. What are the basic combination functions bcfi(..), i =1, .., m, and how do they support the design of network models for any application?\n3. How do mathematical functions and relations enable the analysis of adaptive dynamical systems and self-modeling network models in mental processes?", "prev_section_summary": "The section discusses the mathematical formula for simulating the dynamics of a network model with connectivity, aggregation, and timing characteristics. It also describes the basic combination functions used in network-oriented modeling and how they enable the design of network models. The network model described in (Treur, 2016a) incorporates connectivity, aggregation, and timing characteristics to simulate and analyze network dynamics. The difference (or differential) equations that are useful for simulation and analysis of network dynamics are also discussed.", "section_summary": "The section discusses the use of heterogeneous multi-modal graph neural networks in designing network models and their dynamics in a declarative manner. The basic combination functions bcfi(..), i =1, .., m are used to support the design of network models for any application. Mathematical functions and relations are used to analyze adaptive dynamical systems and self-modeling network models in mental processes. The section also mentions the importance of understanding the states from which incoming connections are obtained.", "excerpt_keywords": "1. Network models\n2. Declarative design\n3. Mathematical functions\n4. Relations\n5. Basic combination functions\n6. J. Treur\n7. Application\n8. Library\n9. Utility\n10. Positivity"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "74571bde-6fcb-4af9-ab8a-35c53b325152", "node_type": "4", "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "d2baa62743b351ba26cd4e7fdac22c4c69777d50fd09182b343bd5c24b7c7286"}, "2": {"node_id": "29fc01a6-e599-43ea-bcc1-2787e7698c86", "node_type": "1", "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "4f5ad3bbbe332963d4a269072a738a8323c8f22f04e1c04538ef070c1d363f9c"}}, "hash": "09157afd1a3f5fbf1b308ff680aea36d887805195e44f86b4efdcf29ff93d707", "text": "are the states from which it gets \nits incoming connections. The above concepts enable to design network \nmodels and their dynamics in a declarative manner, based on mathe -\nmatically defined functions and relations. \nTo support the design of network models, for any application from a \nlibrary predefined basic combination functions bcfi(..), i =1, .., m are J. Treur", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "572359a5-bce2-4e26-9bdb-9369ef6e0c84": {"__data__": {"id_": "572359a5-bce2-4e26-9bdb-9369ef6e0c84", "embedding": null, "metadata": {"page_label": "95", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the mathematical formula used to combine the output of multiple combination functions in a network model?\n2. How are the network characteristics of a model specified in the role matrices format?\n3. What is the difference between adaptive and non-adaptive network models, and how do they interact with each other?", "prev_section_summary": "The section discusses the use of heterogeneous multi-modal graph neural networks in designing network models and their dynamics in a declarative manner. The basic combination functions bcfi(..), i =1, .., m are used to support the design of network models for any application. Mathematical functions and relations are used to analyze adaptive dynamical systems and self-modeling network models in mental processes. The section also mentions the importance of understanding the states from which incoming connections are obtained.", "section_summary": "The section discusses the mathematical formula used to combine the output of multiple combination functions in a network model, the role matrices format used to specify the network characteristics of a model, and the difference between adaptive and non-adaptive network models. The section also explains how self-models can be used within a network to provide a network-oriented conceptualization of the dynamics of the network.", "excerpt_keywords": "1. Network model, 2. Network characteristics, 3. Combination function, 4. Role matrices format, 5. Connectivity characteristics, 6. Aggregation characteristics, 7. Timing characteristics, 8. Adaptive networks, 9. Co-evolution, 10. Self-models"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "1deea7fa-6cef-41a1-96fe-bcf0ad499d43", "node_type": "4", "metadata": {"page_label": "95", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "55541680862a5c09edfb82df46672d43898976f892f04b52241020c46998f67e"}, "3": {"node_id": "93b1b873-30ec-4f04-b951-d4dafe32d0c2", "node_type": "1", "metadata": {"page_label": "95", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "6eea4aa343c0c5169579f08a755873fe349c7109f019a2449b207395d8812a15"}}, "hash": "bff8b55b39c96448d8341c8beb438fa19fb674bdb6f6f93886aacfb8277900e4", "text": "Cognitive Systems Research 70 (2021) 93\u2013100\n95selected by assigning combination function weights \u03b3i,Y, where the com-\nbination function then becomes the weighted average \ncY(.)=\u03b31,Ybcf 1(..)+...+\u03b3m,Ybcf m(..)\n\u03b31,Y+...+\u03b3m,Y(4) \nFurthermore, combination function parameters are specified, so that \nbcfi(..) =bcfi(p,v) where p is a list of parameters (sometimes denoted by \n\u03c0i,j,Y) and v is a list of values. \nA network model is fully defined by its network characteristics. A \nstandardised (and computer-readable) form of specification in table \nformat is used to specify the network characteristics of a model. This \nformat is called the role matrices format; role matrices group the different \ntypes of characteristics:  \n\u2022connectivity characteristics role matrices  \no mb for base connectivity  \no mcw for connection weights \u03c9X,Y  \n\u2022aggregation characteristics role matrices  \no mcfw for combination function weights \u03b3i,Y  \no mcfp for combination function parameters \u03c0i,j,Y  \n\u2022timing characteristics role matrix  \no ms for speed factors \u03b7Y \nExamples of specifications of network models in this standard role \nmatrices format will be shown in Sections 5.2 and 6.2. \nRealistic network models are usually adaptive: their network char-\nacteristics often are adapted over time. Therefore, their dynamics is \nusually an interaction (sometimes called co-evolution) of these two sorts \nof dynamics: dynamics of the nodes (or states) in the network (dynamics within the network) versus dynamics of the characteristics of the \nnetwork (dynamics of the network). By using self-models within the \nnetwork, a network-oriented conceptualisation can also", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "93b1b873-30ec-4f04-b951-d4dafe32d0c2": {"__data__": {"id_": "93b1b873-30ec-4f04-b951-d4dafe32d0c2", "embedding": null, "metadata": {"page_label": "95", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the difference between a self-modeling network model and a traditional network model in terms of their ability to model adaptive networks?\n2. How does the addition of self-model states to a network model provide a declarative description of adaptive network characteristics?\n3. Can self-modeling network models be used to model plasticity and meta-plasticity in the form of a second-order adaptive network with three levels?", "prev_section_summary": "The section discusses the mathematical formula used to combine the output of multiple combination functions in a network model, the role matrices format used to specify the network characteristics of a model, and the difference between adaptive and non-adaptive network models. The section also explains how self-models can be used within a network to provide a network-oriented conceptualization of the dynamics of the network.", "section_summary": "The section discusses the use of self-modeling network models to model adaptive networks. Self-modeling network models are an extension of traditional network models that incorporate new nodes called self-model states to represent adaptive network characteristics. These self-model states are depicted at a next level, called the self-model level, and are connected to nodes at the base level through interlevel connections. The construction of self-modeling network models can be applied iteratively to obtain multiple self-model levels that can provide higher-order adaptive networks. The section provides examples of how self-modeling network models can be used to model plasticity and meta-plasticity in the form of a second-order adaptive network with three levels.", "excerpt_keywords": "1. Network models\n2. Adaptive networks\n3. Self-modeling networks\n4. Dynamics of nodes and network characteristics\n5. Interlevel connections\n6. Mathematical declarative modeling\n7. Plasticity\n8. Meta-plasticity\n9. Higher-order adaptive networks\n10. Self-model states"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "1deea7fa-6cef-41a1-96fe-bcf0ad499d43", "node_type": "4", "metadata": {"page_label": "95", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "55541680862a5c09edfb82df46672d43898976f892f04b52241020c46998f67e"}, "2": {"node_id": "572359a5-bce2-4e26-9bdb-9369ef6e0c84", "node_type": "1", "metadata": {"page_label": "95", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "bff8b55b39c96448d8341c8beb438fa19fb674bdb6f6f93886aacfb8277900e4"}, "3": {"node_id": "96066384-e0d4-45fc-a5d7-3ff5958ec1fb", "node_type": "1", "metadata": {"page_label": "95", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "fae9e8e3df301095aef41c725ee9537a68b0653cae7813e5432488c6324f889c"}}, "hash": "6eea4aa343c0c5169579f08a755873fe349c7109f019a2449b207395d8812a15", "text": "in this standard role \nmatrices format will be shown in Sections 5.2 and 6.2. \nRealistic network models are usually adaptive: their network char-\nacteristics often are adapted over time. Therefore, their dynamics is \nusually an interaction (sometimes called co-evolution) of these two sorts \nof dynamics: dynamics of the nodes (or states) in the network (dynamics within the network) versus dynamics of the characteristics of the \nnetwork (dynamics of the network). By using self-models within the \nnetwork, a network-oriented conceptualisation can also be applied to \nadaptive networks to obtain a declarative description using mathemati -\ncally defined functions and relations; see (Treur, 2020a; Treur, 2020b ). \nThis works through the addition of new nodes to the network (called self- \nmodel states or reification states ) which represent (adaptive) network \ncharacteristics. Such nodes are depicted at a next level (self-model level), \nwhere the original network is at a base level. These types of character -\nistics with their self-model states and their roles are shown in Table 1. \nThis provides an extended network, also called self-modeling network . \nLike for all network models, a self-modeling network model is specified \nin a (network-oriented) declarative mathematical manner based on \nnodes and connections. These include interlevel connections relating \nnodes at one level to nodes on the other. The outcome is also a network \nmodel; so, this whole construction can be applied iteratively to obtain \nmultiple self-model levels that can provide higher-order adaptive net-\nworks, and is quite useful to model, for example, plasticity and meta-\nplasticity in the form of a second-order adaptive network with three \nlevels, one base level and a first- and a second-order self-model level; e. \ng., (Treur, 2020a ) or", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "96066384-e0d4-45fc-a5d7-3ff5958ec1fb": {"__data__": {"id_": "96066384-e0d4-45fc-a5d7-3ff5958ec1fb", "embedding": null, "metadata": {"page_label": "95", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of constructing self-modeling network models in the context of cognitive and neurological modeling?\n2. How can sets of first-order differential equations be used to model causal relationships in various scientific disciplines?\n3. What is the relationship between dynamical systems and network models in the context of computational modeling?", "prev_section_summary": "The section discusses the use of self-modeling network models to model adaptive networks. Self-modeling network models are an extension of traditional network models that incorporate new nodes called self-model states to represent adaptive network characteristics. These self-model states are depicted at a next level, called the self-model level, and are connected to nodes at the base level through interlevel connections. The construction of self-modeling network models can be applied iteratively to obtain multiple self-model levels that can provide higher-order adaptive networks. The section provides examples of how self-modeling network models can be used to model plasticity and meta-plasticity in the form of a second-order adaptive network with three levels.", "section_summary": "The section discusses the construction of self-modeling network models in the context of cognitive and neurological modeling, and their relationship to dynamical systems and network models. Sets of first-order differential equations are used to model causal relationships in various scientific disciplines, including cognitive and neurological modeling. The section also explains how these equations can be used to model plasticity and meta-plasticity in the form of a second-order adaptive network with three levels. The section concludes by discussing the relationship between dynamical systems and network models in computational modeling.", "excerpt_keywords": "1. Network models\n2. Self-model levels\n3. Adaptive networks\n4. Plasticity\n5. Meta-plasticity\n6. Differential equations\n7. Causal relationships\n8. Cognitive modeling\n9. Neurological modeling\n10. Continuous state-determined systems"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "1deea7fa-6cef-41a1-96fe-bcf0ad499d43", "node_type": "4", "metadata": {"page_label": "95", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "55541680862a5c09edfb82df46672d43898976f892f04b52241020c46998f67e"}, "2": {"node_id": "93b1b873-30ec-4f04-b951-d4dafe32d0c2", "node_type": "1", "metadata": {"page_label": "95", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "6eea4aa343c0c5169579f08a755873fe349c7109f019a2449b207395d8812a15"}, "3": {"node_id": "4482903a-bf89-422e-be1f-cffd4bf551fe", "node_type": "1", "metadata": {"page_label": "95", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "19826f3fd7f2b0765cafa57d03c379457d2c4c76332081361beb6e5bc0cda9b2"}}, "hash": "fae9e8e3df301095aef41c725ee9537a68b0653cae7813e5432488c6324f889c", "text": "connections. These include interlevel connections relating \nnodes at one level to nodes on the other. The outcome is also a network \nmodel; so, this whole construction can be applied iteratively to obtain \nmultiple self-model levels that can provide higher-order adaptive net-\nworks, and is quite useful to model, for example, plasticity and meta-\nplasticity in the form of a second-order adaptive network with three \nlevels, one base level and a first- and a second-order self-model level; e. \ng., (Treur, 2020a ) or (Treur, 2020b ), Ch 4. \n5.Relating dynamical systems to network models \nSets of first-order differential equations form a very general format \nused in computational modeling in many disciplines. For cognitive and \nneurological modeling in particular, often causal relationships are used \nin explaining mental processes. But also in many other domains, in a \nwide variety of scientific disciplines causal relationships play a crucial Box 1 \nWhy a smooth continuous state-determined system can be represented by a set of first-order differential equations.  \nSuppose a smooth continuous state-determined system is given. A sketch of why it can be described by a set of first-order differential \nequations is as follows. For any given time point t the future states Xi(t +s) at some future time point time t +s purely depend on s and the \nstates Xi(t) at t. This can be described by (smooth) mathematical functions Fi(\u2026): \nXi(t+s)=Fi(X1(t),\u22ef,Xn(t),s)for s \u22650.  \nIn the particular case of s =0 it holds. \nXi(t)=Fi(X1(t),\u22ef,Xn(t),0)\nSubtracting these two expressions above and dividing by s>0 provides:", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "4482903a-bf89-422e-be1f-cffd4bf551fe": {"__data__": {"id_": "4482903a-bf89-422e-be1f-cffd4bf551fe", "embedding": null, "metadata": {"page_label": "95", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the mathematical function Fi(\u2026) used to describe the dynamics of a heterogeneous multi-modal graph neural network for corporate relative valuation?\n2. How does the function fi(X1, \u2026, Xn) relate to the first-order differential equations that describe the given state-determined system?\n3. What are the different network characteristics and self-model states that can be used to describe the dynamics of a heterogeneous multi-modal graph neural network for corporate relative valuation?", "prev_section_summary": "The section discusses the construction of self-modeling network models in the context of cognitive and neurological modeling, and their relationship to dynamical systems and network models. Sets of first-order differential equations are used to model causal relationships in various scientific disciplines, including cognitive and neurological modeling. The section also explains how these equations can be used to model plasticity and meta-plasticity in the form of a second-order adaptive network with three levels. The section concludes by discussing the relationship between dynamical systems and network models in computational modeling.", "section_summary": "The section discusses the mathematical function Fi(\u2026) used to describe the dynamics of a heterogeneous multi-modal graph neural network for corporate relative valuation. The function relates to the first-order differential equations that describe the given state-determined system. The section also provides a table summarizing the different network characteristics and self-model states that can be used to describe the dynamics of the network.", "excerpt_keywords": "1. Self-model states\n2. Network characteristics\n3. Connectivity characteristics\n4. Aggregation characteristics\n5. First-order differential equations\n6. Mathematical functions\n7. Limits\n8. Differential equations\n9. State-determined system\n10. Network analysis"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "1deea7fa-6cef-41a1-96fe-bcf0ad499d43", "node_type": "4", "metadata": {"page_label": "95", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "55541680862a5c09edfb82df46672d43898976f892f04b52241020c46998f67e"}, "2": {"node_id": "96066384-e0d4-45fc-a5d7-3ff5958ec1fb", "node_type": "1", "metadata": {"page_label": "95", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "fae9e8e3df301095aef41c725ee9537a68b0653cae7813e5432488c6324f889c"}, "3": {"node_id": "fd9baaf3-ca64-4a04-a18d-2e4d634048a2", "node_type": "1", "metadata": {"page_label": "95", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "0a019b8d2b96ee22634406129e35a3f740485245a372f7ed8b170df6a3535619"}}, "hash": "19826f3fd7f2b0765cafa57d03c379457d2c4c76332081361beb6e5bc0cda9b2", "text": "at some future time point time t +s purely depend on s and the \nstates Xi(t) at t. This can be described by (smooth) mathematical functions Fi(\u2026): \nXi(t+s)=Fi(X1(t),\u22ef,Xn(t),s)for s \u22650.  \nIn the particular case of s =0 it holds. \nXi(t)=Fi(X1(t),\u22ef,Xn(t),0)\nSubtracting these two expressions above and dividing by s>0 provides: \nXi(t+s)\u0000Xi(t)\ns=Fi(X1(t),\u22ef,Xn(t),s)\u0000Fi(X1(t),\u22ef,Xn(t),0)\ns  When the limit for s very small, approaching 0 is taken, it follows that: \ndXi(t)\ndt=[\u2202Fi(X1(t),\u22ef,Xn(t),s)\n\u2202s]s=0  \nNow define the function fi(X1, \u2026, Xn) by: \nfi(X1,\u22ef,Xn)=[\u2202Fi(X1(t),\u22ef,Xn(t),s)\n\u2202s]s=0  \nThen it holds. \ndXi(t)\ndt=fi(X1(t),\u22ef,Xn(t))\nThis shows that the given state-determined system can be described by a set of first-order differential equations.    \nTable 1 \nDifferent network characteristics and self-model states for them.  \nTypes of characteristics Concepts Notations Self-model \nstates Role played by the self-model state \nConnectivity \ncharacteristics Connections weights \u03c9X,Y WX,Y Connection weight W \nAggregation characteristics Combination functions weights and \nparameters", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "fd9baaf3-ca64-4a04-a18d-2e4d634048a2": {"__data__": {"id_": "fd9baaf3-ca64-4a04-a18d-2e4d634048a2", "embedding": null, "metadata": {"page_label": "95", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. How does the use of first-order differential equations help in describing the state-determined system in the given context?\n2. What are the different types of network characteristics and self-model states discussed in Table 1, and how do they play a role in the system?\n3. How do the combination functions, weights, and parameters affect the behavior of the system, and what is their significance in the context of self-modeling network models?", "prev_section_summary": "The section discusses the mathematical function Fi(\u2026) used to describe the dynamics of a heterogeneous multi-modal graph neural network for corporate relative valuation. The function relates to the first-order differential equations that describe the given state-determined system. The section also provides a table summarizing the different network characteristics and self-model states that can be used to describe the dynamics of the network.", "section_summary": "The section discusses the use of first-order differential equations to describe the state-determined system in a given context. It also introduces different types of network characteristics and self-model states, and their roles in the system. The combination functions, weights, and parameters are also discussed and their significance in the context of self-modeling network models is highlighted. The section concludes with a summary of the key topics and entities.", "excerpt_keywords": "1. Network characteristics\n2. Self-model states\n3. Connectivity\n4. Aggregation\n5. Timing\n6. Speed factors\n7. Weight\n8. Parameter\n9. Combination function\n10. Utility"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "1deea7fa-6cef-41a1-96fe-bcf0ad499d43", "node_type": "4", "metadata": {"page_label": "95", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "55541680862a5c09edfb82df46672d43898976f892f04b52241020c46998f67e"}, "2": {"node_id": "4482903a-bf89-422e-be1f-cffd4bf551fe", "node_type": "1", "metadata": {"page_label": "95", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "19826f3fd7f2b0765cafa57d03c379457d2c4c76332081361beb6e5bc0cda9b2"}}, "hash": "0a019b8d2b96ee22634406129e35a3f740485245a372f7ed8b170df6a3535619", "text": "shows that the given state-determined system can be described by a set of first-order differential equations.    \nTable 1 \nDifferent network characteristics and self-model states for them.  \nTypes of characteristics Concepts Notations Self-model \nstates Role played by the self-model state \nConnectivity \ncharacteristics Connections weights \u03c9X,Y WX,Y Connection weight W \nAggregation characteristics Combination functions weights and \nparameters \u03b3i,Y \u03c0i,j,Y Ci,Y Pi,j,Y Combination function weight C Combination function parameter \nP \nTiming characteristics Speed factors \u03b7Y HY Speed factor H  J. Treur", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "c5fc4b53-c961-43ee-93a3-a051764fb0c3": {"__data__": {"id_": "c5fc4b53-c961-43ee-93a3-a051764fb0c3", "embedding": null, "metadata": {"page_label": "96", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. How can a state-determined system be transformed into a format that more directly relates to causal relationships between states?\n2. What is the conceptual representation of a dynamical system as a causal graph, and how can it be visualized?\n3. How can a numerical representation of a dynamical system be transformed into a causal graph model, and what is the mathematical formula for this transformation?", "prev_section_summary": "The section discusses the use of first-order differential equations to describe the state-determined system in a given context. It also introduces different types of network characteristics and self-model states, and their roles in the system. The combination functions, weights, and parameters are also discussed and their significance in the context of self-modeling network models is highlighted. The section concludes with a summary of the key topics and entities.", "section_summary": "The section discusses the transformation of state-determined systems into a format that more directly relates to causal relationships between states. This can be achieved by representing the dynamical system as a causal graph, which can be visualized in a network model. The section also provides a mathematical formula for transforming a numerical representation of a dynamical system into a causal graph model. The key topics and entities discussed in the section include dynamical systems, causal relationships, graph neural networks, and differential equations.", "excerpt_keywords": "1. Dynamical systems\n2. Network models\n3. Causal relationships\n4. Graphical representation\n5. Temporal-causal network format\n6. State-determined systems\n7. State changes\n8. Subset of states\n9. Causal effects\n10. Differential equations"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "5acd044b-9d8d-47f7-b88e-cf4ea4514f96", "node_type": "4", "metadata": {"page_label": "96", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "bceeaf74bbe465a1db4d83091b131127e70915875abf2743d5784551523b8ddb"}, "3": {"node_id": "c7751fc1-42be-4936-81e6-4344d4449a51", "node_type": "1", "metadata": {"page_label": "96", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "4b7e18b4b03ac594e2eafae429753d61997e4817f582ac75a5b3eeffb994084b"}}, "hash": "0d32f8ac05493f5f7943d880735e10c5b1a4f2b763788497c3c003a2e90a2288", "text": "Cognitive Systems Research 70 (2021) 93\u2013100\n96role. In this context it will be useful if it can be explained more explicitly \nhow any state-determined system can be described or transformed into a \nformat that more directly relates to causal relationships between states. \nThis indeed can always be achieved in the temporal-causal network \nformat (when arbitrary combination functions are allowed), in the \nmanner shown below in general in Section 5.1 and illustrated by an \nexample in Section 5.2. \n5.1. Transforming a dynamical system model into a network model \nIn a dynamical system the changes in each state S depend on the \nother states. Those states R that actually play a role in this dependence \nrelation form a subset DS of the set of all states (in some special cases this \nsubset may be the set of all states). The states R not in this subset DS are \nthose states for which never any state change of R has influence on a \nchange of the state of S. The states in this subset DS can be considered to \ncause the changes in the state S. Such causal effects of states on each \nother by causal relationships can be visualized in a graphical manner as, \nfor example, in the network shown in Fig. 3. Such a network model can \nbe simply defined by the following criterion: for any state R and any \nstate S there is a connection from R to S if and only if R \u2208DS. This \nprovides a conceptual representation of the dynamical system as a \ncausal graph. \nAnother transformation can be done for the numerical representation \non the basis of a set of differential equations representing the dynamical \nsystem. Suppose a differential equation for one of the states Xi is given of \nthe form: \ndXi(t)\ndt=fi(X1(t),\u22ef,Xn(t)) (5) \nThen this function fi(X1(t), \u2026,", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "c7751fc1-42be-4936-81e6-4344d4449a51": {"__data__": {"id_": "c7751fc1-42be-4936-81e6-4344d4449a51", "embedding": null, "metadata": {"page_label": "96", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the conceptual representation of a dynamical system as a causal graph, and how does it relate to the numerical representation of the system using differential equations?\n2. How can the function fi(X1(t), \u2026, Xn(t)) be defined for a given differential equation of the form dXi(t) dt=fi(X1(t), \u2026, Xn(t)), and what is the subset DXi of the set of states {X1, \u2026, Xn} that actually occur in this formula?\n3. What is the difference between the differential equation (5) for Xi and the differential equation (7) for Xi, and how is the function hi(X1, \u2026, Xn) defined in relation to these equations?", "prev_section_summary": "The section discusses the transformation of state-determined systems into a format that more directly relates to causal relationships between states. This can be achieved by representing the dynamical system as a causal graph, which can be visualized in a network model. The section also provides a mathematical formula for transforming a numerical representation of a dynamical system into a causal graph model. The key topics and entities discussed in the section include dynamical systems, causal relationships, graph neural networks, and differential equations.", "section_summary": "The section discusses the conceptual representation of a dynamical system as a causal graph and the numerical representation of the system using differential equations. It explains how the function fi(X1(t), \u2026, Xn(t)) can be defined for a given differential equation of the form dXi(t) dt=fi(X1(t), \u2026, Xn(t)), and how the subset DXi of the set of states {X1, \u2026, Xn} that actually occur in this formula can be defined. The section also explains the difference between the differential equation (5) for Xi and the differential equation (7) for Xi, and how the function hi(X1, \u2026, Xn) is defined in relation to these equations.", "excerpt_keywords": "1. Dynamical system, 2. Causal graph, 3. Numerical representation, 4. Differential equations, 5. State variables, 6. Dependency relations, 7. Subset, 8. Formula, 9. Causal connection, 10. Functional representation"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "5acd044b-9d8d-47f7-b88e-cf4ea4514f96", "node_type": "4", "metadata": {"page_label": "96", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "bceeaf74bbe465a1db4d83091b131127e70915875abf2743d5784551523b8ddb"}, "2": {"node_id": "c5fc4b53-c961-43ee-93a3-a051764fb0c3", "node_type": "1", "metadata": {"page_label": "96", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "0d32f8ac05493f5f7943d880735e10c5b1a4f2b763788497c3c003a2e90a2288"}, "3": {"node_id": "0a646399-e0de-4836-b25d-04ae232449c3", "node_type": "1", "metadata": {"page_label": "96", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "2721aefb2ea148d524ebdb5c331b8924285d38999980af3676b4b89d8efe57c8"}}, "hash": "4b7e18b4b03ac594e2eafae429753d61997e4817f582ac75a5b3eeffb994084b", "text": "is a connection from R to S if and only if R \u2208DS. This \nprovides a conceptual representation of the dynamical system as a \ncausal graph. \nAnother transformation can be done for the numerical representation \non the basis of a set of differential equations representing the dynamical \nsystem. Suppose a differential equation for one of the states Xi is given of \nthe form: \ndXi(t)\ndt=fi(X1(t),\u22ef,Xn(t)) (5) \nThen this function fi(X1(t), \u2026, Xn(t)) will depend on a subset DXi of \nthe set of states {X1, \u2026, Xn}. Note that Xi itself may occur in DXi. Usually \nthis function fi will be given as a formula in X1, \u2026, Xn; then this subset \ncan be taken as the set of all states in {X1, \u2026, Xn} that actually occur in \nthis formula. Again, for any two states Xj and Xi with j \u2215=i a causal \nconnection from Xj to Xi can be defined by the criterion that Xj \u2208DXi. \nNow, by defining the function hi(X1, \u2026, Xn) by \nhi(X1,\u22ef,Xn)=Xi+fi(X1,\u22ef,Xn) (6) \ndifferential equation (5) for Xi always can be rewritten into a dif-\nferential equation of the form \ndXi(t)\ndt=\u03b7i[ci(\u03c91,iX1(t),\u22ef,\u03c9n,iXn(t))\u0000Xi(t)] (7) \nwhere \u03b7i =1, and also \u03c9j,i =1, and for combination function ci(\u2026) it \nholds ci(\u2026) =hi(\u2026). Note that Xi itself may occur in ci(\u2026). So, having", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "0a646399-e0de-4836-b25d-04ae232449c3": {"__data__": {"id_": "0a646399-e0de-4836-b25d-04ae232449c3", "embedding": null, "metadata": {"page_label": "96", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the mathematical representation of a temporal-causal network model for a continuous smooth dynamical system?\n2. Can any type of combination function be used to describe a temporal-causal network model for a continuous smooth dynamical system?\n3. What is the relationship between temporal-causal network models and dynamical system models?", "prev_section_summary": "The section discusses the conceptual representation of a dynamical system as a causal graph and the numerical representation of the system using differential equations. It explains how the function fi(X1(t), \u2026, Xn(t)) can be defined for a given differential equation of the form dXi(t) dt=fi(X1(t), \u2026, Xn(t)), and how the subset DXi of the set of states {X1, \u2026, Xn} that actually occur in this formula can be defined. The section also explains the difference between the differential equation (5) for Xi and the differential equation (7) for Xi, and how the function hi(X1, \u2026, Xn) is defined in relation to these equations.", "section_summary": "The section discusses the relationship between temporal-causal network models and dynamical system models. It shows that any continuous smooth dynamical system can be described by a specific temporal-causal network model, as long as any type of combination function is allowed. This relationship is illustrated by an example in Section 5.2. The section also provides a mathematical representation of a temporal-causal network model for a continuous smooth dynamical system.", "excerpt_keywords": "dynamical system, network model, differential equation, temporal-causal network, continuous, smooth, combination function, representation, transformation, illustration, example."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "5acd044b-9d8d-47f7-b88e-cf4ea4514f96", "node_type": "4", "metadata": {"page_label": "96", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "bceeaf74bbe465a1db4d83091b131127e70915875abf2743d5784551523b8ddb"}, "2": {"node_id": "c7751fc1-42be-4936-81e6-4344d4449a51", "node_type": "1", "metadata": {"page_label": "96", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "4b7e18b4b03ac594e2eafae429753d61997e4817f582ac75a5b3eeffb994084b"}, "3": {"node_id": "07a0d979-1c2f-46c4-a0a7-91de295cfd65", "node_type": "1", "metadata": {"page_label": "96", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "f9668cceb3f9fce780248018d7cdd7b012af30c4d2d7b9f683c43c04f33c57bc"}}, "hash": "2721aefb2ea148d524ebdb5c331b8924285d38999980af3676b4b89d8efe57c8", "text": "differential equation (5) for Xi always can be rewritten into a dif-\nferential equation of the form \ndXi(t)\ndt=\u03b7i[ci(\u03c91,iX1(t),\u22ef,\u03c9n,iXn(t))\u0000Xi(t)] (7) \nwhere \u03b7i =1, and also \u03c9j,i =1, and for combination function ci(\u2026) it \nholds ci(\u2026) =hi(\u2026). Note that Xi itself may occur in ci(\u2026). So, having \nstarted with any arbitrary continuous, smooth dynamical system and its \nrepresentation (5) in differential equation format, finally a numerical \nrepresentation of a temporal-causal network model was obtained. This \nshows that any continuous smooth dynamical system can be described \nby a specific temporal-causal network model, as long as any type of \ncombination function is allowed. So, the following theorem has been \nobtained: \nTheorem 1 (from dynamical system to network model) \nAny continuous smooth dynamical system model can be transformed \ninto a temporal-causal network model. Conversely, any temporal-causal \nnetwork model is a dynamical system model. \nThis Theorem 1 and the underlying transformation will be illustrated \nby an example in Section 5.2. \n5.2. Illustration of the transformation for an example dynamical system \nAs an illustration, consider an arbitrary example of a dynamical \nsystem model described by numerical first-order differential equations \nrepresentation with parameters \u03b1, \u03b2, \u03b3, \u03b4:", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "07a0d979-1c2f-46c4-a0a7-91de295cfd65": {"__data__": {"id_": "07a0d979-1c2f-46c4-a0a7-91de295cfd65", "embedding": null, "metadata": {"page_label": "96", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the relationship between continuous smooth dynamical system models and temporal-causal network models?\n2. How can an arbitrary example of a dynamical system model be transformed into a temporal-causal network model?\n3. What are the connections between the states in the example dynamical system model (8) when transformed into a temporal-causal network model?", "prev_section_summary": "The section discusses the relationship between temporal-causal network models and dynamical system models. It shows that any continuous smooth dynamical system can be described by a specific temporal-causal network model, as long as any type of combination function is allowed. This relationship is illustrated by an example in Section 5.2. The section also provides a mathematical representation of a temporal-causal network model for a continuous smooth dynamical system.", "section_summary": "The section discusses the relationship between continuous smooth dynamical system models and temporal-causal network models. It explains that any continuous smooth dynamical system model can be transformed into a temporal-causal network model, and vice versa. The section provides an example of how to transform an arbitrary example of a dynamical system model into a temporal-causal network model. The example is represented in a graphical network format as shown in Fig. 1. The section also mentions that the connections between the states in the example dynamical system model (8) when transformed into a temporal-causal network model are X5 affects X1, X1 and X3 affect X2, X2 affects X3, X3 and X5 affect X4, and X4 affects X5.", "excerpt_keywords": "dynamical systems, temporal-causal network model, transformation, first-order differential equations, parameters, states, connections, graphical network format, illustration, example."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "5acd044b-9d8d-47f7-b88e-cf4ea4514f96", "node_type": "4", "metadata": {"page_label": "96", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "bceeaf74bbe465a1db4d83091b131127e70915875abf2743d5784551523b8ddb"}, "2": {"node_id": "0a646399-e0de-4836-b25d-04ae232449c3", "node_type": "1", "metadata": {"page_label": "96", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "2721aefb2ea148d524ebdb5c331b8924285d38999980af3676b4b89d8efe57c8"}, "3": {"node_id": "e85dffe1-a000-42dd-99a7-4c5eabf7ef14", "node_type": "1", "metadata": {"page_label": "96", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "e1abf177dbf75c97fdb4440840c71e9f92470c14a3f8a96b6c0e666da7b02566"}}, "hash": "f9668cceb3f9fce780248018d7cdd7b012af30c4d2d7b9f683c43c04f33c57bc", "text": "to network model) \nAny continuous smooth dynamical system model can be transformed \ninto a temporal-causal network model. Conversely, any temporal-causal \nnetwork model is a dynamical system model. \nThis Theorem 1 and the underlying transformation will be illustrated \nby an example in Section 5.2. \n5.2. Illustration of the transformation for an example dynamical system \nAs an illustration, consider an arbitrary example of a dynamical \nsystem model described by numerical first-order differential equations \nrepresentation with parameters \u03b1, \u03b2, \u03b3, \u03b4: dX1(t)\ndt=X1(t)(X5(t)\u0000\u03b1)\ndX2(t)\ndt=X1(t)\u0000X2(t)+X3(t)\ndX3(t)\ndt=X2(t)(\u03b2\u0000X3(t))\ndX4(t)\ndt=X3(t)\u0000X4(t)(\u03b3\u0000X5(t))\ndX5(t)\ndt=X5(t)(\u03b4\u0000X4(t)) (8) \nTo transform this dynamical system model (8) into a temporal-causal \nnetwork model, the five states X1, X2, X3, X4, X5 are considered. From \neach of the equations by inspecting which states occur in the right hand \nside it can subsequently be determined that:  \nX5 affects X1 \nX1 and X3 affect X2 \nX2 affects X3 \nX3 and X5 affect X4 \nX4 affects X5  \nThese connections are represented in a graphical network format as \nshown in Fig. 1. \nNote that, when comparing, for example, the first differential equa-\ntion in (8) to the standard format of differential equations for temporal- \ncausal networks, it can be written as", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "e85dffe1-a000-42dd-99a7-4c5eabf7ef14": {"__data__": {"id_": "e85dffe1-a000-42dd-99a7-4c5eabf7ef14", "embedding": null, "metadata": {"page_label": "96", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the mathematical analysis used to determine the connections between variables in the graphical network format shown in Fig. 1?\n2. How does the combination function cX1(..) defined in equation (10) affect the differential equation dX1(t) dt=X1(t)(X5(t) - \u03b1) +X1(t)?\n3. What is the relationship between the combination functions cX1(..), cX2(..), and cX3(..) and the differential equations (8)?", "prev_section_summary": "The section discusses the relationship between continuous smooth dynamical system models and temporal-causal network models. It explains that any continuous smooth dynamical system model can be transformed into a temporal-causal network model, and vice versa. The section provides an example of how to transform an arbitrary example of a dynamical system model into a temporal-causal network model. The example is represented in a graphical network format as shown in Fig. 1. The section also mentions that the connections between the states in the example dynamical system model (8) when transformed into a temporal-causal network model are X5 affects X1, X1 and X3 affect X2, X2 affects X3, X3 and X5 affect X4, and X4 affects X5.", "section_summary": "The section discusses the mathematical analysis used to determine the connections between variables in a graphical network format, specifically in relation to adaptive dynamical systems and self-modeling network models. The combination function cX1(..) is defined and its relationship with the differential equations (8) is explained. The section also identifies other combination functions (cX2(..) and cX3(..)) built from sum and product functions, and their relationship with the differential equations (8) is determined. The key topics and entities discussed in the section include graphical network format, adaptive dynamical systems, self-modeling network models, mathematical analysis, combination functions, and differential equations.", "excerpt_keywords": "1. Differential equations, \n2. Temporal-causal networks, \n3. Graphical network format, \n4. Combination functions, \n5. Sum and product functions, \n6. Causality, \n7. Positive and fairness, \n8. Utility, \n9. Ethical considerations, \n10. Prejudiced content."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "5acd044b-9d8d-47f7-b88e-cf4ea4514f96", "node_type": "4", "metadata": {"page_label": "96", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "bceeaf74bbe465a1db4d83091b131127e70915875abf2743d5784551523b8ddb"}, "2": {"node_id": "07a0d979-1c2f-46c4-a0a7-91de295cfd65", "node_type": "1", "metadata": {"page_label": "96", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "f9668cceb3f9fce780248018d7cdd7b012af30c4d2d7b9f683c43c04f33c57bc"}, "3": {"node_id": "e9e11adf-a9e5-42a1-9eed-7b53e2289646", "node_type": "1", "metadata": {"page_label": "96", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "6943be33afa823bcdb89b6ece2a9e82e766340743c3e9c7686f819ac5943dd17"}}, "hash": "e1abf177dbf75c97fdb4440840c71e9f92470c14a3f8a96b6c0e666da7b02566", "text": "the equations by inspecting which states occur in the right hand \nside it can subsequently be determined that:  \nX5 affects X1 \nX1 and X3 affect X2 \nX2 affects X3 \nX3 and X5 affect X4 \nX4 affects X5  \nThese connections are represented in a graphical network format as \nshown in Fig. 1. \nNote that, when comparing, for example, the first differential equa-\ntion in (8) to the standard format of differential equations for temporal- \ncausal networks, it can be written as \ndX1(t)\ndt=X1(t)(X5(t)\u0000\u03b1)\n=X1(t)(X5(t)\u0000\u03b1)+X1(t)\u0000X1(t)\n=[X1(t)(X5(t)\u0000\u03b1)+X1(t)]\u0000X1(t) (9) \nHere the part (X1(t) (X5(t) - \u03b1) +X1(t)) can be considered the result \nof a combination function cX1(..) defined by \ncX1(V1,V5)=V1+V1(V5\u0000\u03b1) (10) \napplied to X1(t) (for V1) and X5(t) (for V5). In a similar manner the \nfollowing combination functions (built from sum and product functions) \ncan be identified from the five differential equations (8): \ncX1(V1,V5)=V1+V1(V5\u0000\u03b1)=(1\u0000\u03b1)V1+V1V5 (11)  \ncX2(V1,V3)=V2+V1\u0000V2+V3=V1+V3  \ncX3(V2,V3)=V3+V2(\u03b2\u0000V3)=\u03b2V2+V3\u0000V2V3", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "e9e11adf-a9e5-42a1-9eed-7b53e2289646": {"__data__": {"id_": "e9e11adf-a9e5-42a1-9eed-7b53e2289646", "embedding": null, "metadata": {"page_label": "96", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the mathematical representation of the combination functions used in the given differential equations?\n2. How do the combination functions relate to the graphical network representation of the example model?\n3. What is the significance of the combination functions in the context of adaptive dynamical systems and self-modeling network models?", "prev_section_summary": "The section discusses the mathematical analysis used to determine the connections between variables in a graphical network format, specifically in relation to adaptive dynamical systems and self-modeling network models. The combination function cX1(..) is defined and its relationship with the differential equations (8) is explained. The section also identifies other combination functions (cX2(..) and cX3(..)) built from sum and product functions, and their relationship with the differential equations (8) is determined. The key topics and entities discussed in the section include graphical network format, adaptive dynamical systems, self-modeling network models, mathematical analysis, combination functions, and differential equations.", "section_summary": "The section discusses the mathematical representation of combination functions used in differential equations related to adaptive dynamical systems and self-modeling network models. The combination functions are built from sum and product functions and are used to model the dynamics and adaptivity of mental processes. The section also provides a graphical network representation of the example model based on the given differential equation representation. The significance of the combination functions in the context of these models is not explicitly stated.", "excerpt_keywords": "1. Differential equations\n2. Graphical network representation\n3. Combination functions\n4. Utility\n5. Fairness\n6. Positivity\n7. Care\n8. Respect\n9. Truth\n10. Modeling"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "5acd044b-9d8d-47f7-b88e-cf4ea4514f96", "node_type": "4", "metadata": {"page_label": "96", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "bceeaf74bbe465a1db4d83091b131127e70915875abf2743d5784551523b8ddb"}, "2": {"node_id": "e85dffe1-a000-42dd-99a7-4c5eabf7ef14", "node_type": "1", "metadata": {"page_label": "96", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "e1abf177dbf75c97fdb4440840c71e9f92470c14a3f8a96b6c0e666da7b02566"}}, "hash": "6943be33afa823bcdb89b6ece2a9e82e766340743c3e9c7686f819ac5943dd17", "text": "combination functions (built from sum and product functions) \ncan be identified from the five differential equations (8): \ncX1(V1,V5)=V1+V1(V5\u0000\u03b1)=(1\u0000\u03b1)V1+V1V5 (11)  \ncX2(V1,V3)=V2+V1\u0000V2+V3=V1+V3  \ncX3(V2,V3)=V3+V2(\u03b2\u0000V3)=\u03b2V2+V3\u0000V2V3  \ncX4(V3,V4,V5)=V4+V3\u0000V4(\u03b3\u0000V5)=V3+(1\u0000\u03b3)V4+V4V5  \ncX5(V4,V5)=V5+V5(\u03b4\u0000V4)=(1+\u03b4)V5\u0000V4V5 \nUsing these combination functions, the differential equations (8) can \nFig. 1.Graphical network representation for the example model based on the \ngiven differential equation representation. J. Treur", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "67141005-9c62-489d-9374-bff1fa9da7ae": {"__data__": {"id_": "67141005-9c62-489d-9374-bff1fa9da7ae", "embedding": null, "metadata": {"page_label": "97", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the mathematical relationship between adaptive dynamical systems and self-modeling network models?\n2. How can the temporal-causal network model described by the differential equations (8) be transformed into an instance of a more general temporal-causal model?\n3. What is the significance of the numerical representation (3) in the context of temporal-causal network models?", "prev_section_summary": "The section discusses the mathematical representation of combination functions used in differential equations related to adaptive dynamical systems and self-modeling network models. The combination functions are built from sum and product functions and are used to model the dynamics and adaptivity of mental processes. The section also provides a graphical network representation of the example model based on the given differential equation representation. The significance of the combination functions in the context of these models is not explicitly stated.", "section_summary": "The section discusses the relationship between adaptive dynamical systems and self-modeling network models, specifically in the context of temporal-causal network models. The section presents a numerical representation of a temporal-causal network model and explains how the differential equations representation can be transformed into an instance of the numerical representation of a more general temporal-causal model. The section also discusses the significance of the numerical representation in the context of temporal-causal network models.", "excerpt_keywords": "1. Temporal-causal network model\n2. Dynamical system model\n3. Numerical representation\n4. Differential equations\n5. Transformation\n6. Exact manner\n7. General temporal-causal model\n8. \u03b7Xi =1\n9. \u03c9XiXj =1\n10. Cognitive Systems Research"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "1a293364-9a65-4363-a78e-033057f53944", "node_type": "4", "metadata": {"page_label": "97", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "687e845953e6f61280551d93d2040d81948690f1932acc74aac2afa24f3c62fa"}, "3": {"node_id": "59b16652-6ade-45b3-9ad5-6bbaebd78823", "node_type": "1", "metadata": {"page_label": "97", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "ba4a4a70a71e5842bdc4ca8a68ec95353e37393b06069320a95c570f1110a583"}}, "hash": "27f98466bd354bb37217a806be518c5fa3c0835459e61ae2dca8ea5e29cde52a", "text": "Cognitive Systems Research 70 (2021) 93\u2013100\n97be rewritten into: \ndX1(t)\ndt=cX1(X1(t),X5(t))\u0000X1(t)\ndX2(t)\ndt=cX2(X1(t),X3(t))\u0000X1(t)\ndX3(t)\ndt=cX3(X2(t),X3(t))\u0000X3(t)\ndX4(t)\ndt=cX4(X3(t),X4(t),X5(t))\u0000X4(t)\ndX5(t)\ndt=cX5(X4(t),X5(t))\u0000X5(t) (12) \nThis is the numerical representation (3) of a temporal-causal network \nmodel with \u03b7Xi =1 for all i and \u03c9XiXj =1 for all i and j. It turns out that \nthe dynamical system model described by the differential equations \nrepresentation (8) can be transformed in an exact manner into an \ninstance of the numerical representation of the more general temporal- \ncausal model described by the following differential equations:", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "59b16652-6ade-45b3-9ad5-6bbaebd78823": {"__data__": {"id_": "59b16652-6ade-45b3-9ad5-6bbaebd78823", "embedding": null, "metadata": {"page_label": "97", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the mathematical representation of a temporal-causal network model with \u03b7Xi =1 for all i and \u03c9XiXj =1 for all i and j?\n2. How can the dynamical system model described by the differential equations representation (8) be transformed into an instance of the numerical representation of a more general temporal-causal model?\n3. What is the role matrix representation of the corresponding network model based on the given set of differential equations, and what are the parameter values used in the representation?", "prev_section_summary": "The section discusses the relationship between adaptive dynamical systems and self-modeling network models, specifically in the context of temporal-causal network models. The section presents a numerical representation of a temporal-causal network model and explains how the differential equations representation can be transformed into an instance of the numerical representation of a more general temporal-causal model. The section also discusses the significance of the numerical representation in the context of temporal-causal network models.", "section_summary": "The section discusses the mathematical representation of a temporal-causal network model with \u03b7Xi =1 for all i and \u03c9XiXj =1 for all i and j. It also explains how the dynamical system model described by the differential equations representation (8) can be transformed into an instance of the numerical representation of a more general temporal-causal model. The section provides the role matrix representation of the corresponding network model based on the given set of differential equations, with parameter values of 1 chosen.", "excerpt_keywords": "1. Temporal-causal network model\n2. Differential equations representation\n3. Numerical representation\n4. Role matrix representation\n5. Parameter values\n6. State Xj\n7. Network characteristics\n8. \u03c9X,Y, cY, \u03b7Y\n9. Exact transformation\n10. Positive and fairness-promoting replies"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "1a293364-9a65-4363-a78e-033057f53944", "node_type": "4", "metadata": {"page_label": "97", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "687e845953e6f61280551d93d2040d81948690f1932acc74aac2afa24f3c62fa"}, "2": {"node_id": "67141005-9c62-489d-9374-bff1fa9da7ae", "node_type": "1", "metadata": {"page_label": "97", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "27f98466bd354bb37217a806be518c5fa3c0835459e61ae2dca8ea5e29cde52a"}, "3": {"node_id": "9e946097-f0e9-444f-8e38-8402b039005c", "node_type": "1", "metadata": {"page_label": "97", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "5a935b4368920ba63df09df4d1d0bd314b4c71e2c9766f5bf6d06f861d1bf37e"}}, "hash": "ba4a4a70a71e5842bdc4ca8a68ec95353e37393b06069320a95c570f1110a583", "text": "(12) \nThis is the numerical representation (3) of a temporal-causal network \nmodel with \u03b7Xi =1 for all i and \u03c9XiXj =1 for all i and j. It turns out that \nthe dynamical system model described by the differential equations \nrepresentation (8) can be transformed in an exact manner into an \ninstance of the numerical representation of the more general temporal- \ncausal model described by the following differential equations: \ndX1(t)\ndt=\u03b7X1[cX1\u0000\n(\u03c9X1,X1X1(t),\u03c9X5,X1X5(t))\n\u0000X1(t)]\ndX2(t)\ndt=\u03b7X2[cX2\u0000\u03c9X1,X2X1(t),\u03c9X3,X2X3(t))\n\u0000X2(t)]\ndX3(t)\ndt=\u03b7X3[cX3\u0000\u03c9X2,X3X2(t),\u03c9X3,X3X3(t))\n\u0000X3(t)]\ndX4(t)\ndt=\u03b7X4[cX4\u0000\u03c9X3,X4X3(t),\u03c9X4,X4X4(t),\u03c9X5,X4X5(t))\n\u0000X4(t)\ndX5(t)\ndt=\u03b7X5[cX5\u0000\u03c9X4,X5X4(t),\u03c9X5,X5X5(t))\n\u0000X5(t)] (13) \nBased on these network characteristics \u03c9X,Y, cY, \u03b7Y found from the \ngiven set of differential equations, the role matrix representation of the \ncorresponding network model (with parameter values 1 chosen) can be \nobtained as shown in Fig. 2. Here the different rows in each matrix \naddress the different states Xj of", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "9e946097-f0e9-444f-8e38-8402b039005c": {"__data__": {"id_": "9e946097-f0e9-444f-8e38-8402b039005c", "embedding": null, "metadata": {"page_label": "97", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the role matrix representation in the given network model, and how does it help in understanding the connectivity, aggregation, and timing characteristics of the network?\n2. How do the connection weights \u03c9Xi,Xj specified in the role matrix mcw affect the behavior of the network, and what is their significance in the context of the given differential equations?\n3. What is the role of the combination functions and their parameters in the network model, and how do they contribute to the overall behavior of the network?", "prev_section_summary": "The section discusses the mathematical representation of a temporal-causal network model with \u03b7Xi =1 for all i and \u03c9XiXj =1 for all i and j. It also explains how the dynamical system model described by the differential equations representation (8) can be transformed into an instance of the numerical representation of a more general temporal-causal model. The section provides the role matrix representation of the corresponding network model based on the given set of differential equations, with parameter values of 1 chosen.", "section_summary": "The section discusses the role matrix representation in a network model and its significance in understanding the connectivity, aggregation, and timing characteristics of the network. The connection weights \u03c9Xi,Xj specified in the role matrix mcw affect the behavior of the network and are significant in the context of the given differential equations. The combination functions and their parameters in the network model contribute to the overall behavior of the network. The section also explains how the role matrix representation of the corresponding network model can be obtained based on the network characteristics found from the given set of differential equations.", "excerpt_keywords": "1. Network characteristics\n2. Role matrix representation\n3. Differential equations\n4. Connectivity characteristics\n5. Aggregation characteristics\n6. Timing characteristics\n7. Parameter values\n8. Connection weights\n9. Combination functions\n10. State transitions"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "1a293364-9a65-4363-a78e-033057f53944", "node_type": "4", "metadata": {"page_label": "97", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "687e845953e6f61280551d93d2040d81948690f1932acc74aac2afa24f3c62fa"}, "2": {"node_id": "59b16652-6ade-45b3-9ad5-6bbaebd78823", "node_type": "1", "metadata": {"page_label": "97", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "ba4a4a70a71e5842bdc4ca8a68ec95353e37393b06069320a95c570f1110a583"}, "3": {"node_id": "894b0c9d-4aac-46ea-9b88-5450d1f33057", "node_type": "1", "metadata": {"page_label": "97", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "3ef265b7016febbd86c5def6707ceaa709c9a70990ea8224a04f129bce3b3c6c"}}, "hash": "5a935b4368920ba63df09df4d1d0bd314b4c71e2c9766f5bf6d06f861d1bf37e", "text": "(13) \nBased on these network characteristics \u03c9X,Y, cY, \u03b7Y found from the \ngiven set of differential equations, the role matrix representation of the \ncorresponding network model (with parameter values 1 chosen) can be \nobtained as shown in Fig. 2. Here the different rows in each matrix \naddress the different states Xj of the network model. Recall that they \ngroup the different types of characteristics \nConnectivity characteristics  \u2022In role matrix mb at each row for the indicated state Xj it is specified \nfrom which other states it gets incoming connections. For example, \nin the second row it is indicated that state X2 has incoming connec -\ntions from state X1 and X3.  \n\u2022In role matrix mcw at each row for the indicated state Xj the \nconnection weight \u03c9Xi,Xj is specified for each of the incoming con-\nnections for Xj as indicated in mb. For example, in the second row it \nis indicated that state X2 has connection weights 1 for both incoming \nconnections (as all other states have). \nAggregation characteristics  \n\u2022In role matrix mcfw at each row for the indicated state Xj it is \nspecified which combination function(s) are used and with what \nweight(s) \u03b3i,Y. For example, in the fourth row it is indicated that state \nX4 uses combination function cX4(..) with weight 1.  \n\u2022In role matrix mcfp at each row for the indicated state Xj the \nparameter values \u03c0i,j,Xi of the combination function(s) are specified \nfor each of the combination functions indicated in mcfw . For \nexample, in the fourth row it is indicated that the combination \nfunction indicated in mcfw for state X4 uses value 1 for parameter \u03b3. \nTiming characteristics  \n\u2022In role matrix ms at each row for the indicated state", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "894b0c9d-4aac-46ea-9b88-5450d1f33057": {"__data__": {"id_": "894b0c9d-4aac-46ea-9b88-5450d1f33057", "embedding": null, "metadata": {"page_label": "97", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using combination functions and weight values in the role matrix mcfp for the adaptive dynamical system model?\n2. How does the timing characteristics specified in the role matrix ms affect the performance of the adaptive dynamical system model?\n3. Can you explain how the approach described in Section 5 can be extended to obtain a transformation of an adaptive dynamical system into a self-modeling network model, and provide an example of how this transformation can be illustrated for an adaptive dynamical system?", "prev_section_summary": "The section discusses the role matrix representation in a network model and its significance in understanding the connectivity, aggregation, and timing characteristics of the network. The connection weights \u03c9Xi,Xj specified in the role matrix mcw affect the behavior of the network and are significant in the context of the given differential equations. The combination functions and their parameters in the network model contribute to the overall behavior of the network. The section also explains how the role matrix representation of the corresponding network model can be obtained based on the network characteristics found from the given set of differential equations.", "section_summary": "The section discusses the use of combination functions and weight values in the role matrix mcfp for the adaptive dynamical system model, and how timing characteristics specified in the role matrix ms affect the performance of the adaptive dynamical system model. It also describes how the approach described in Section 5 can be extended to obtain a transformation of an adaptive dynamical system into a self-modeling network model, and provides an example of how this transformation can be illustrated for an adaptive dynamical system. The section includes a figure showing the general approach and another figure illustrating the transformation for an example adaptive dynamical system.", "excerpt_keywords": "1. Adaptive dynamical systems\n2. Self-modeling network models\n3. Role matrices\n4. Combination functions\n5. Timing characteristics\n6. Speed factors\n7. Differential equations\n8. Network models\n9. Transformation\n10. Example set"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "1a293364-9a65-4363-a78e-033057f53944", "node_type": "4", "metadata": {"page_label": "97", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "687e845953e6f61280551d93d2040d81948690f1932acc74aac2afa24f3c62fa"}, "2": {"node_id": "9e946097-f0e9-444f-8e38-8402b039005c", "node_type": "1", "metadata": {"page_label": "97", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "5a935b4368920ba63df09df4d1d0bd314b4c71e2c9766f5bf6d06f861d1bf37e"}, "3": {"node_id": "1d88f0c5-ed50-41e1-99f4-b6bc66e49c64", "node_type": "1", "metadata": {"page_label": "97", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "bd5e11acb5888d0d5cc8a108e7a9f3b109a362022de70d4a77d3463423cd10b4"}}, "hash": "3ef265b7016febbd86c5def6707ceaa709c9a70990ea8224a04f129bce3b3c6c", "text": "state \nX4 uses combination function cX4(..) with weight 1.  \n\u2022In role matrix mcfp at each row for the indicated state Xj the \nparameter values \u03c0i,j,Xi of the combination function(s) are specified \nfor each of the combination functions indicated in mcfw . For \nexample, in the fourth row it is indicated that the combination \nfunction indicated in mcfw for state X4 uses value 1 for parameter \u03b3. \nTiming characteristics  \n\u2022In role matrix ms at each row for the indicated state Xj it is specified \nwhich speed factor is used. It shows that for this network model all \nspeed factors are 1. \n6.Relating adaptive dynamical systems to self-modeling \nnetwork models \nIn this section, it is shown how the approach described in Section 5 \ncan be extended to obtain a transformation of an adaptive dynamical \nsystem into a self-modeling network model. In Section 6.1 the general \napproach is described and in Section 6.2 it is illustrated for an example \nadaptive dynamical system. \n6.1. Transforming an adaptive dynamical system model into a self- \nmodeling network model \nAdaptive dynamical systems are usually modeled by two levels of \ndynamical systems (see Fig. 3) where the higher level dynamical system \nFig. 2.Role matrices specification of the example set of differential equations.  J. Treur", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "1d88f0c5-ed50-41e1-99f4-b6bc66e49c64": {"__data__": {"id_": "1d88f0c5-ed50-41e1-99f4-b6bc66e49c64", "embedding": null, "metadata": {"page_label": "97", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the relationship between adaptive dynamical systems and self-modeling network models in mental processes?\n2. How can mathematical analysis be used to understand the dynamics and adaptivity of mental processes?\n3. What is the role of role matrices in specifying the example set of differential equations in the context of mental processes?", "prev_section_summary": "The section discusses the use of combination functions and weight values in the role matrix mcfp for the adaptive dynamical system model, and how timing characteristics specified in the role matrix ms affect the performance of the adaptive dynamical system model. It also describes how the approach described in Section 5 can be extended to obtain a transformation of an adaptive dynamical system into a self-modeling network model, and provides an example of how this transformation can be illustrated for an adaptive dynamical system. The section includes a figure showing the general approach and another figure illustrating the transformation for an example adaptive dynamical system.", "section_summary": "The section discusses the relationship between adaptive dynamical systems and self-modeling network models in mental processes, and how mathematical analysis can be used to understand the dynamics and adaptivity of mental processes. The role of role matrices in specifying the example set of differential equations in the context of mental processes is also discussed. The section is authored by J. Treur.", "excerpt_keywords": "1. Differential equations\n2. Role matrices\n3. Specification\n4. Example set\n5. Utility\n6. Security\n7. Harmful content\n8. Unethical content\n9. Prejudiced content\n10. Positivity"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "1a293364-9a65-4363-a78e-033057f53944", "node_type": "4", "metadata": {"page_label": "97", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "687e845953e6f61280551d93d2040d81948690f1932acc74aac2afa24f3c62fa"}, "2": {"node_id": "894b0c9d-4aac-46ea-9b88-5450d1f33057", "node_type": "1", "metadata": {"page_label": "97", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "3ef265b7016febbd86c5def6707ceaa709c9a70990ea8224a04f129bce3b3c6c"}}, "hash": "bd5e11acb5888d0d5cc8a108e7a9f3b109a362022de70d4a77d3463423cd10b4", "text": "2.Role matrices specification of the example set of differential equations.  J. Treur", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "cb0ecf56-27a0-45f6-b6ea-9169066c6cb8": {"__data__": {"id_": "cb0ecf56-27a0-45f6-b6ea-9169066c6cb8", "embedding": null, "metadata": {"page_label": "98", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the mathematical model used to describe the dynamics of variables Xi in the lower level dynamical system?\n2. How do the parameters Pi,j become time-dependent in the upper level component of the model?\n3. What is the transformation applied to the base dynamical system (14) to obtain the combination functions hi(\u2026) found for the base level states Xi?", "prev_section_summary": "The section discusses the relationship between adaptive dynamical systems and self-modeling network models in mental processes, and how mathematical analysis can be used to understand the dynamics and adaptivity of mental processes. The role of role matrices in specifying the example set of differential equations in the context of mental processes is also discussed. The section is authored by J. Treur.", "section_summary": "The section discusses the mathematical model used to describe the dynamics of variables Xi in the lower level dynamical system, how the parameters Pi,j become time-dependent in the upper level component of the model, and how the transformation applied to the base dynamical system (14) obtains the combination functions hi(\u2026) found for the base level states Xi. Additionally, the section describes how a similar transformation is applied to the adaptation level described by (15) to define function qi,j(..). The key topics and entities discussed in this section include adaptive dynamical systems, self-modeling network models, mathematical analysis, heterogeneous multi-modal graph neural networks, and corporate relative valuation.", "excerpt_keywords": "1. Dynamical systems\n2. Parameters\n3. Lower level component\n4. Upper level component\n5. Time-dependent\n6. Differential equations\n7. Transformation\n8. Combination functions\n9. Adaptation level\n10. Pi,j"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "b0d761ef-9bec-46bf-a112-261d0f389c3a", "node_type": "4", "metadata": {"page_label": "98", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a9e4c033eccff1f4e383d17e78f9f752da3a5aaa165f444af85d153aff779acf"}, "3": {"node_id": "39a7d3e7-ab49-4d23-8cef-4bd4f8a5b3b2", "node_type": "1", "metadata": {"page_label": "98", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "aa09ffc529f2d3ed161824e4e1331a53b4a4597965f66765dc217785eb67341f"}}, "hash": "3c96ebfbb84421a250d93b4a4f92c417d6ca697300c8dfe6b704c9c1522b1625", "text": "Cognitive Systems Research 70 (2021) 93\u2013100\n98models the dynamics of the parameters Pi,j of the lower level dynamical \nsystem (the lower level component in Fig. 3) that describes the dynamics \nof variables Xi, for example by \ndXi(t)\ndt=fi\u0000\nPi,1,\u22ef,Pi,k,X1(t),\u22ef,Xn(t))\n(14) \nThese parameters Pi,j then become time-dependent: Pi,j(t). In addi-\ntion, for the dynamics of the Pi,j there will also be differential equations \n(the upper level component in Fig. 3): \ndPi,j(t)\ndt=pi,j\u0000\nP1,1(t),\u22ef,Pn,k(t),X1(t),\u22ef,Xn(t))\n(15) \nApplying the transformation described in Section 5 on the base \ndynamical system (14) for the Xi, these Pi,j will also become parameters \nof the combination functions hi(\u2026) found for the base level states Xi: \nhi\u0000\nPi,1,\u22ef,Pi,k,X1,\u22ef,Xn)\n=Xi+fi\u0000\nPi,1,\u22ef,Pi,k,X1,\u22ef,Xn)\n(16) \nApplying a similar transformation to the adaptation level described \nby (15), in this case define function qi,j(..) by:   \nThen", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "39a7d3e7-ab49-4d23-8cef-4bd4f8a5b3b2": {"__data__": {"id_": "39a7d3e7-ab49-4d23-8cef-4bd4f8a5b3b2", "embedding": null, "metadata": {"page_label": "98", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the mathematical formula used to model the combination of base level states Xi and Pi,j in a heterogeneous multi-modal graph neural network for corporate relative valuation?\n2. How can the adaptation level described by (15) be modeled at the first-order self-model level as self-model states for the base level?\n3. What is the theorem that states that any adaptive continuous smooth dynamical system model can be transformed into a self-modeling temporal-causal network model, and vice versa?", "prev_section_summary": "The section discusses the mathematical model used to describe the dynamics of variables Xi in the lower level dynamical system, how the parameters Pi,j become time-dependent in the upper level component of the model, and how the transformation applied to the base dynamical system (14) obtains the combination functions hi(\u2026) found for the base level states Xi. Additionally, the section describes how a similar transformation is applied to the adaptation level described by (15) to define function qi,j(..). The key topics and entities discussed in this section include adaptive dynamical systems, self-modeling network models, mathematical analysis, heterogeneous multi-modal graph neural networks, and corporate relative valuation.", "section_summary": "The section discusses the relationship between adaptive dynamical systems and self-modeling network models in the context of modeling corporate relative valuation. The mathematical formula used to model the combination of base level states Xi and Pi,j in a heterogeneous multi-modal graph neural network for corporate relative valuation is given in equation (16). The adaptation level described by (15) is modeled at the first-order self-model level as self-model states for the base level using equation (18). The theorem that states that any adaptive continuous smooth dynamical system model can be transformed into a self-modeling temporal-causal network model, and vice versa, is presented as Theorem 2.", "excerpt_keywords": "1. Adaptive dynamical system\n2. Self-modeling network model\n3. Continuous smooth dynamical system\n4. Temporal-causal network model\n5. First-order self-modeling\n6. Higher-order self-modeling\n7. Higher-order adaptive dynamical system\n8. Combination function\n9. Connection weights\n10. Transformation"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "b0d761ef-9bec-46bf-a112-261d0f389c3a", "node_type": "4", "metadata": {"page_label": "98", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a9e4c033eccff1f4e383d17e78f9f752da3a5aaa165f444af85d153aff779acf"}, "2": {"node_id": "cb0ecf56-27a0-45f6-b6ea-9169066c6cb8", "node_type": "1", "metadata": {"page_label": "98", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "3c96ebfbb84421a250d93b4a4f92c417d6ca697300c8dfe6b704c9c1522b1625"}, "3": {"node_id": "e41dabeb-3e80-45bf-b38f-785a14c25b34", "node_type": "1", "metadata": {"page_label": "98", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "d3a097eb66004504bf3902493c427dabbb81932ba1972cc61d7cee6f2fc89f22"}}, "hash": "aa09ffc529f2d3ed161824e4e1331a53b4a4597965f66765dc217785eb67341f", "text": "for the Xi, these Pi,j will also become parameters \nof the combination functions hi(\u2026) found for the base level states Xi: \nhi\u0000\nPi,1,\u22ef,Pi,k,X1,\u22ef,Xn)\n=Xi+fi\u0000\nPi,1,\u22ef,Pi,k,X1,\u22ef,Xn)\n(16) \nApplying a similar transformation to the adaptation level described \nby (15), in this case define function qi,j(..) by:   \nThen \ndPi,j(t)\ndt=\u03b7i,j[ci,j(\n\u03c91,1,i,jP1,1(t),\u22ef,\u03c9n,k,i,jPn,k(t),\u03c91,i,jX1(t),\u22ef,\u03c9n,i,jXn(t))\n\u0000Pi,j(t)]\n(18) \nwhere again \u03b7i,j =1 and all connection weights \u03c9 involved are 1 too \nand for the combination function ci,j(\u2026), it simply holds ci,j(\u2026) =qi, \nj(\u2026). By this, the states Pi,j can be modeled at the first-order self-model \nlevel as self-model states for the base level. In a similar manner it can be \nshown by iteration that any higher-order adaptive dynamical system can \nbe modelled as a higher-order self-modeling network model. Thus the \nfollowing theorem is obtained: \nTheorem 2 (from adaptive dynamical system to self-modeling \nnetwork model) \nAny adaptive continuous smooth dynamical system model can be \ntransformed into a self-modeling temporal-causal network model. \nConversely, any self-modeling temporal-causal network model is an \nadaptive dynamical system model. These also apply to higher-order", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "e41dabeb-3e80-45bf-b38f-785a14c25b34": {"__data__": {"id_": "e41dabeb-3e80-45bf-b38f-785a14c25b34", "embedding": null, "metadata": {"page_label": "98", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the relationship between adaptive dynamical systems and self-modeling network models in terms of mathematical analysis?\n2. Can any higher-order adaptive dynamical system be modeled as a higher-order self-modeling network model?\n3. What is the underlying transformation between adaptive dynamical systems and self-modeling network models, and how can it be illustrated with an example?", "prev_section_summary": "The section discusses the relationship between adaptive dynamical systems and self-modeling network models in the context of modeling corporate relative valuation. The mathematical formula used to model the combination of base level states Xi and Pi,j in a heterogeneous multi-modal graph neural network for corporate relative valuation is given in equation (16). The adaptation level described by (15) is modeled at the first-order self-model level as self-model states for the base level using equation (18). The theorem that states that any adaptive continuous smooth dynamical system model can be transformed into a self-modeling temporal-causal network model, and vice versa, is presented as Theorem 2.", "section_summary": "The section discusses the relationship between adaptive dynamical systems and self-modeling network models in terms of mathematical analysis. It explains that any adaptive continuous smooth dynamical system model can be transformed into a self-modeling temporal-causal network model, and vice versa. The transformation is illustrated by an example of an adaptive extension of a previously described system, where the four parameters are made adaptive and differential equations are assumed for the Pi,j states. The section also mentions that this transformation applies to higher-order adaptive dynamical systems in relation to higher-order self-modeling networks.", "excerpt_keywords": "Adaptive dynamical systems, self-modeling network models, higher-order systems, continuous smooth dynamical systems, temporal-causal network models, differential equations, Pi,j states, X1, X2, X3, X4, X5, X6, X7, X8, X9."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "b0d761ef-9bec-46bf-a112-261d0f389c3a", "node_type": "4", "metadata": {"page_label": "98", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a9e4c033eccff1f4e383d17e78f9f752da3a5aaa165f444af85d153aff779acf"}, "2": {"node_id": "39a7d3e7-ab49-4d23-8cef-4bd4f8a5b3b2", "node_type": "1", "metadata": {"page_label": "98", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "aa09ffc529f2d3ed161824e4e1331a53b4a4597965f66765dc217785eb67341f"}, "3": {"node_id": "f331d817-57db-4bd2-aedb-d960b44cf61e", "node_type": "1", "metadata": {"page_label": "98", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "3de0938c1d709a1d287650a43f1601a613127713a6c049d566204f6b8907d47d"}}, "hash": "d3a097eb66004504bf3902493c427dabbb81932ba1972cc61d7cee6f2fc89f22", "text": "can be \nshown by iteration that any higher-order adaptive dynamical system can \nbe modelled as a higher-order self-modeling network model. Thus the \nfollowing theorem is obtained: \nTheorem 2 (from adaptive dynamical system to self-modeling \nnetwork model) \nAny adaptive continuous smooth dynamical system model can be \ntransformed into a self-modeling temporal-causal network model. \nConversely, any self-modeling temporal-causal network model is an \nadaptive dynamical system model. These also apply to higher-order \nadaptive dynamical systems in relation to higher-order self-modeling \nnetworks. \nThis Theorem 2 and the underlying transformation will be illustrated \nby an example in Section 6.2. 6.2. Illustration of the transformation for an example adaptive dynamical \nsystem \nThe transformation described in Section 6.1 can be applied to an \nadaptive extension of the example described in Section 5.2 by making \nthe four parameters \u03b1, \u03b2, \u03b3, \u03b4 adaptive. To this end, for each of these \nparameters \u03b1, \u03b2, \u03b3, \u03b4 a differential equation is assumed for the Pi,j states \nused, where: \nP1,1=\u03b1P3,1=\u03b2P4,1=\u03b3P5,1=\u03b4 (19) \nExtending the example, suppose these are the equations for them: \ndP1,1(t)\ndt=X1(t)+X5(t)\ndP3,1(t)\ndt=X2(t)+X3(t)\ndP4,1(t)\ndt=X3(t)+X4(t)\ndP5,1(t)\ndt=X4(t)+X5(t) (20) \nNow rename them as self-model states X6, X7, X8, X9, respectively:", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "f331d817-57db-4bd2-aedb-d960b44cf61e": {"__data__": {"id_": "f331d817-57db-4bd2-aedb-d960b44cf61e", "embedding": null, "metadata": {"page_label": "98", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the relationship between adaptive dynamical systems and self-modeling network models in the context of mental processes?\n2. How can mathematical analysis be used to understand the dynamics and adaptivity of mental processes?\n3. What are the implications of using heterogeneous multi-modal graph neural networks for corporate relative valuation?", "prev_section_summary": "The section discusses the relationship between adaptive dynamical systems and self-modeling network models in terms of mathematical analysis. It explains that any adaptive continuous smooth dynamical system model can be transformed into a self-modeling temporal-causal network model, and vice versa. The transformation is illustrated by an example of an adaptive extension of a previously described system, where the four parameters are made adaptive and differential equations are assumed for the Pi,j states. The section also mentions that this transformation applies to higher-order adaptive dynamical systems in relation to higher-order self-modeling networks.", "section_summary": "The section discusses the relationship between adaptive dynamical systems and self-modeling network models in the context of mental processes. Mathematical analysis is used to understand the dynamics and adaptivity of mental processes. The use of heterogeneous multi-modal graph neural networks for corporate relative valuation is also discussed, with the equations (20) being rewritten in the standard network format using combination functions named cX6 to cX9. The new combination functions are defined by adaptation and base dynamical system parameters.", "excerpt_keywords": "1. Self-model states\n2. Network self-model\n3. Combination functions\n4. Adaptation system\n5. Base dynamical system\n6. Parameter\n7. Equations\n8. Rename\n9. Standard network format\n10. Utility"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "b0d761ef-9bec-46bf-a112-261d0f389c3a", "node_type": "4", "metadata": {"page_label": "98", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a9e4c033eccff1f4e383d17e78f9f752da3a5aaa165f444af85d153aff779acf"}, "2": {"node_id": "e41dabeb-3e80-45bf-b38f-785a14c25b34", "node_type": "1", "metadata": {"page_label": "98", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "d3a097eb66004504bf3902493c427dabbb81932ba1972cc61d7cee6f2fc89f22"}, "3": {"node_id": "ce244400-592a-41ac-9992-aa96623a8a49", "node_type": "1", "metadata": {"page_label": "98", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "7464148b739e3fb04a7b221ac723b04f23265d46fc625d6559739b5cf20c196f"}}, "hash": "3de0938c1d709a1d287650a43f1601a613127713a6c049d566204f6b8907d47d", "text": "the example, suppose these are the equations for them: \ndP1,1(t)\ndt=X1(t)+X5(t)\ndP3,1(t)\ndt=X2(t)+X3(t)\ndP4,1(t)\ndt=X3(t)+X4(t)\ndP5,1(t)\ndt=X4(t)+X5(t) (20) \nNow rename them as self-model states X6, X7, X8, X9, respectively: \nX6=P1,1=\u03b1X7=P3,1=\u03b2X8=P4,1=\u03b3X9=P5,1=\u03b4 (21) \nThen the above equations (20) for the parameters (now denoted by \nadditional network self-model states X6 to X9) can be rewritten (as in \nSection 6.1) in the standard network format using combination func-\ntions named cX6, cX7, cX8, cX9, respectively, leading to the equations \ndX6(t)\ndt=cX6(X1(t),X5(t),X6(t))\u0000X6(t)\ndX7(t)\ndt=cX7(X2(t),X3(t),X7(t))\u0000X7(t)\ndX8(t)\ndt=cX8(X3(t),X4(t),X8(t))\u0000X8(t)\ndX9(t)\ndt=cX9(X4(t),X5(t),X9(t))\u0000X9(t) (22) \nwhere the new combination functions cX6(..) to cX9(..) are defined by \nAdaptation  \nsystem\nBase dynamical  \nsystem Parameter", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "ce244400-592a-41ac-9992-aa96623a8a49": {"__data__": {"id_": "ce244400-592a-41ac-9992-aa96623a8a49", "embedding": null, "metadata": {"page_label": "98", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the mathematical analysis of adaptive dynamical systems and self-modeling network models?\n2. How do the new combination functions cX6(..) to cX9(..) relate to the adaptation system and base dynamical system in an adaptive dynamical system?\n3. What is the graphical representation of the self-modeling network model based on the given differential equation representation of an adaptive dynamical system?", "prev_section_summary": "The section discusses the relationship between adaptive dynamical systems and self-modeling network models in the context of mental processes. Mathematical analysis is used to understand the dynamics and adaptivity of mental processes. The use of heterogeneous multi-modal graph neural networks for corporate relative valuation is also discussed, with the equations (20) being rewritten in the standard network format using combination functions named cX6 to cX9. The new combination functions are defined by adaptation and base dynamical system parameters.", "section_summary": "The section discusses the mathematical analysis of adaptive dynamical systems and self-modeling network models. It explains the new combination functions cX6(..) to cX9(..) and their relationship to the adaptation system and base dynamical system in an adaptive dynamical system. The section also provides a graphical representation of the self-modeling network model based on the given differential equation representation of an adaptive dynamical system. The key topics and entities discussed in the section include adaptive dynamical systems, self-modeling network models, mathematical analysis, combination functions, and graphical representation.", "excerpt_keywords": "Adaptive dynamical system, self-modeling network model, differential equation representation, parameter values, input values, Xi, qi, P1, P2, P3, P4, P5, P6, P7, P8, P9, P10, Pi, jValuesXi, X1, X2, X3, X4, X5, X6, X7, X8, X9."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "b0d761ef-9bec-46bf-a112-261d0f389c3a", "node_type": "4", "metadata": {"page_label": "98", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a9e4c033eccff1f4e383d17e78f9f752da3a5aaa165f444af85d153aff779acf"}, "2": {"node_id": "f331d817-57db-4bd2-aedb-d960b44cf61e", "node_type": "1", "metadata": {"page_label": "98", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "3de0938c1d709a1d287650a43f1601a613127713a6c049d566204f6b8907d47d"}}, "hash": "7464148b739e3fb04a7b221ac723b04f23265d46fc625d6559739b5cf20c196f", "text": "(22) \nwhere the new combination functions cX6(..) to cX9(..) are defined by \nAdaptation  \nsystem\nBase dynamical  \nsystem Parameter valuesPi,jValuesXi\nInput \nvaluesXiFig. 3.Architecture of an adaptive dynamical system.  \nX1X5X4\nX2 X3X6 X8X9\nX7\nFig. 4.Graphical representation of the self-modeling network model based on \nthe given differential equation representation of an adaptive dynamical system. qi,j\u0000\nP1,1(t),\u22ef,Pn,k(t),X1(t),\u22ef,Xn(t))\n=Pi,j(t)+pi,j\u0000\nP1,1(t),\u22ef,Pn,k(t),X1(t),\u22ef,Xn(t))\n(17)   J. Treur", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "69ee04f9-a2b5-4f68-a89f-5e0a468ed6bb": {"__data__": {"id_": "69ee04f9-a2b5-4f68-a89f-5e0a468ed6bb", "embedding": null, "metadata": {"page_label": "99", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the mathematical relationship between network-oriented modeling and the dynamical systems perspective on mental processes?\n2. Can any adaptive dynamical system be accurately modeled using a self-modeling network?\n3. What is the philosophical significance of the relationship between dynamical systems and network-oriented modeling in understanding mental processes?", "prev_section_summary": "The section discusses the mathematical analysis of adaptive dynamical systems and self-modeling network models. It explains the new combination functions cX6(..) to cX9(..) and their relationship to the adaptation system and base dynamical system in an adaptive dynamical system. The section also provides a graphical representation of the self-modeling network model based on the given differential equation representation of an adaptive dynamical system. The key topics and entities discussed in the section include adaptive dynamical systems, self-modeling network models, mathematical analysis, combination functions, and graphical representation.", "section_summary": "The section discusses the relationship between network-oriented modeling and the dynamical systems perspective on mental processes. It explains how any dynamical system can be modeled as a temporal-causal network model and any adaptive dynamical system can be modeled by a self-modeling network. The section also provides mathematical transformations for these models and discusses their philosophical significance in understanding mental processes. The key topics and entities discussed in the section include network-oriented modeling, dynamical systems perspective, mental processes, adaptive dynamical systems, temporal-causal network models, self-modeling networks, and mathematical transformations.", "excerpt_keywords": "dynamical systems, mental processes, network-oriented modeling, temporal-causal network model, adaptive dynamical systems, self-modeling network, mathematical transformations, exact modeling, philosophical context, dynamical systems perspective."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "d64f1393-8ea3-4495-b60d-300797f050d1", "node_type": "4", "metadata": {"page_label": "99", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "e4a007ae349c076b957891f562c88e341d8c0e16c5ce8eaf008e24e30dd3c3cb"}, "3": {"node_id": "4fd74f11-08d6-41bd-a62d-3fe64bbaee4a", "node_type": "1", "metadata": {"page_label": "99", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "43993cd17b3fdef9ac91d90a7b16ead6d23d00ca5bce40c63b400fd05a71444f"}}, "hash": "355c4d033afd70d815cf458314f2d975fd4dfa78a59481fe17e3b91bd2ad39df", "text": "Cognitive Systems Research 70 (2021) 93\u2013100\n99cX6(V1,V5,V6)=V1+V5+V6  \ncX7(V2,V3,V7)=V2+V3+V7  \ncX8(V3,V4,V8)=V3+V4+V8  \ncX9(V4,V5,V9)=V4+V5+V9 (23) \nHere Vi are variables used for Xi(t). Note that these four functions \ncX6, cX7, cX8, cX9 are actually the same function, namely the sum func-\ntion; but for generality of the illustration they will be named by their \ndifferent names. This then will get the form of the example picture as \nshown in Fig. 4 and the role matrices in Fig. 5. \n7.Discussion \nIn this paper it was addressed how network-oriented modeling re-\nlates to the dynamical systems perspective on mental processes as \ndescribed by (Ashby, 1960; Beer, 2000; Kelso, 1995; Port and Van \nGelder, 1995; Scherer, 2009; Thelen and Smith, 1994; van Gelder, 1998; \nvan Gelder and Port, 1995 ). It has been mathematically proven that any \ndynamical system can be modeled as a temporal-causal network model \naccording to the approach described in (Treur, 2016a; Treur, 2016b ) \nand that any adaptive dynamical system (of any order) can be modeled \nby a self-modeling network (of the same order) as described in (Treur, \n2020a; Treur, 2020b ). Note that the approach presented here gives exact \nmathematical transformations, not approximations such as, for example \ndescribed in (Funahashi and Nakamura, 1993 ). \nIn a wider philosophical context, the dynamical", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "4fd74f11-08d6-41bd-a62d-3fe64bbaee4a": {"__data__": {"id_": "4fd74f11-08d6-41bd-a62d-3fe64bbaee4a", "embedding": null, "metadata": {"page_label": "99", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the mathematical transformation for modeling any adaptive dynamical system by a self-modeling network according to the approach described in (Treur, 2016a; Treur, 2016b)?\n2. How does the dynamical systems perspective relate to the role of temporal factorisation and criterial causation in adaptive dynamical systems, as described in (Treur, 2007a; Treur, 2007b; Treur, 2021; Tse, 2013)?\n3. What is the philosophical foundation of the dynamical systems perspective, and how does it relate to earlier ideas on dynamics introduced by Descartes and Laplace?", "prev_section_summary": "The section discusses the relationship between network-oriented modeling and the dynamical systems perspective on mental processes. It explains how any dynamical system can be modeled as a temporal-causal network model and any adaptive dynamical system can be modeled by a self-modeling network. The section also provides mathematical transformations for these models and discusses their philosophical significance in understanding mental processes. The key topics and entities discussed in the section include network-oriented modeling, dynamical systems perspective, mental processes, adaptive dynamical systems, temporal-causal network models, self-modeling networks, and mathematical transformations.", "section_summary": "The section discusses the mathematical transformation for modeling any adaptive dynamical system by a self-modeling network according to the approach described in (Treur, 2016a; Treur, 2016b). The dynamical systems perspective is also related to the role of temporal factorisation and criterial causation in adaptive dynamical systems, as described in (Treur, 2007a; Treur, 2007b; Treur, 2021; Tse, 2013). The philosophical foundation of the dynamical systems perspective is the notion of state-determined system (Ashby, 1960) and relates to earlier ideas on dynamics introduced by Descartes (1634) and Laplace (1825). The current paper shows that network-oriented modeling provides adequate means to model such dynamics and adaptivity.", "excerpt_keywords": "1. Temporal-causal network model\n2. Adaptive dynamical systems\n3. Self-modeling network\n4. State-determined system\n5. Temporal factorisation\n6. Criterial causation\n7. Dynamical systems perspective\n8. Descartes' dynamics\n9. Laplace's dynamics\n10. Network-oriented modeling"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "d64f1393-8ea3-4495-b60d-300797f050d1", "node_type": "4", "metadata": {"page_label": "99", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "e4a007ae349c076b957891f562c88e341d8c0e16c5ce8eaf008e24e30dd3c3cb"}, "2": {"node_id": "69ee04f9-a2b5-4f68-a89f-5e0a468ed6bb", "node_type": "1", "metadata": {"page_label": "99", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "355c4d033afd70d815cf458314f2d975fd4dfa78a59481fe17e3b91bd2ad39df"}, "3": {"node_id": "3ee06420-3e90-4e62-bdc6-a81fd84a1271", "node_type": "1", "metadata": {"page_label": "99", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a05214ed12213e7f306c2beacfea066665968eb8036cfc498d30eeabc43b70a9"}}, "hash": "43993cd17b3fdef9ac91d90a7b16ead6d23d00ca5bce40c63b400fd05a71444f", "text": "as a temporal-causal network model \naccording to the approach described in (Treur, 2016a; Treur, 2016b ) \nand that any adaptive dynamical system (of any order) can be modeled \nby a self-modeling network (of the same order) as described in (Treur, \n2020a; Treur, 2020b ). Note that the approach presented here gives exact \nmathematical transformations, not approximations such as, for example \ndescribed in (Funahashi and Nakamura, 1993 ). \nIn a wider philosophical context, the dynamical systems perspective \nhas its foundation in the notion of state-determined system (Ashby, \n1960 ), and relates to the role the notions of temporal factorisation and \ncriterial causation play in adaptive dynamical systems (Treur, 2007a; \nTreur, 2007b; Treur, 2021; Tse, 2013 ) and to the earlier ideas on dy-\nnamics introduced by Descartes (1634) and Laplace (1825) . What the \ncurrent paper shows is that from a mathematical analysis perspective, \nnetwork-oriented modeling provide adequates means to model such \ndynamics and adaptivity. Declaration of Competing Interest \nThe authors declare that they have no known competing financial \ninterests or personal relationships that could have appeared to influence \nthe work reported in this paper. \nReferences \nAshby, W.R.: Design for a Brain. Chapman and Hall, London (second extended edition). \nFirst edition, 1952 (1960). \nBeer, R. D. (2000). Dynamical approaches to cognitive science. Trends in Cognitive \nSciences, 4(3), 91\u201399. \nDescartes, R. (1634) The World Ch 6: Description of a New World and on the Qualities of \nthe Matter of Which it is Composed. \nFunahashi, K.-I., & Nakamura, Y. (1993). Approximation of dynamical", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "3ee06420-3e90-4e62-bdc6-a81fd84a1271": {"__data__": {"id_": "3ee06420-3e90-4e62-bdc6-a81fd84a1271", "embedding": null, "metadata": {"page_label": "99", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the relationship between adaptive dynamical systems and self-modeling network models in the context of mental processes?\n2. How can mathematical analysis be used to understand the dynamics and adaptivity of mental processes?\n3. What are some examples of mathematical models that have been used to study the dynamics and adaptivity of mental processes?", "prev_section_summary": "The section discusses the mathematical transformation for modeling any adaptive dynamical system by a self-modeling network according to the approach described in (Treur, 2016a; Treur, 2016b). The dynamical systems perspective is also related to the role of temporal factorisation and criterial causation in adaptive dynamical systems, as described in (Treur, 2007a; Treur, 2007b; Treur, 2021; Tse, 2013). The philosophical foundation of the dynamical systems perspective is the notion of state-determined system (Ashby, 1960) and relates to earlier ideas on dynamics introduced by Descartes (1634) and Laplace (1825). The current paper shows that network-oriented modeling provides adequate means to model such dynamics and adaptivity.", "section_summary": "The section discusses the relationship between adaptive dynamical systems and self-modeling network models in the context of mental processes. Mathematical analysis can be used to understand the dynamics and adaptivity of mental processes, and examples of mathematical models that have been used to study this include the work of Funahashi and Nakamura, Kuipers and Kassirer, and Port and van Gelder. Other relevant works mentioned include W.R.'s \"Design for a Brain,\" Beer's \"Dynamical approaches to cognitive science,\" Descartes' \"The World,\" Kelso's \"Dynamic patterns,\" Kim's \"Philosophy of Mind,\" Kuipers' \"Commonsense reasoning about causality,\" Laplace's \"Philosophical Essays on Probabilities,\" Pearl's \"Causality,\" and Port and van Gelder's \"Mind as motion.\"", "excerpt_keywords": "1. Brain design\n2. Cognitive science\n3. Dynamical systems\n4. Neural networks\n5. Self-organization\n6. Philosophy of mind\n7. Causality\n8. Commonsense reasoning\n9. Knowledge representation\n10. Probabilities"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "d64f1393-8ea3-4495-b60d-300797f050d1", "node_type": "4", "metadata": {"page_label": "99", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "e4a007ae349c076b957891f562c88e341d8c0e16c5ce8eaf008e24e30dd3c3cb"}, "2": {"node_id": "4fd74f11-08d6-41bd-a62d-3fe64bbaee4a", "node_type": "1", "metadata": {"page_label": "99", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "43993cd17b3fdef9ac91d90a7b16ead6d23d00ca5bce40c63b400fd05a71444f"}, "3": {"node_id": "0b71a2c3-b090-4c63-9246-2c508d56a785", "node_type": "1", "metadata": {"page_label": "99", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "fb85dd74236bf749c9d90b8acbf4fe018886e380c1fc8b28a9ddc1266e90eb0e"}}, "hash": "a05214ed12213e7f306c2beacfea066665968eb8036cfc498d30eeabc43b70a9", "text": "W.R.: Design for a Brain. Chapman and Hall, London (second extended edition). \nFirst edition, 1952 (1960). \nBeer, R. D. (2000). Dynamical approaches to cognitive science. Trends in Cognitive \nSciences, 4(3), 91\u201399. \nDescartes, R. (1634) The World Ch 6: Description of a New World and on the Qualities of \nthe Matter of Which it is Composed. \nFunahashi, K.-I., & Nakamura, Y. (1993). Approximation of dynamical systems by \ncontinuous time recurrent neural networks. Neural Networks, 6(6), 801\u2013806. \nKelso, J. A. S. (1995). Dynamic patterns: The self-organization of brain and behavior . \nCambridge: MIT Press .  \nKim, J. (1996). Philosophy of Mind. Westview Press .  \nKuipers, B. J. (1984). Commonsense reasoning about causality: Deriving behavior from \nstructure. Artificial Intelligence, 24(1984), 169\u2013203. \nKuipers, B.J. and Kassirer, J.P., (1983). How to discover a knowledge representation for \ncausal reasoning by studying an expert physician. In: Proceedings Eighth \nInternational Joint Conference on Artificial Intelligence, IJCAI \u201983, Karlsruhe, F.R.G. \n(William Kaufman, Los Altos, CA, 1983). \nLaplace, P.S. (1825) Philosophical Essays on Probabilities Springer-Verlag New York \n1995 Translated by AI Dale from the 5th French edition of 1825. \nPearl, J. (2000). 2000 . Causality: Cambridge University Press .  \nPort, R. F., & van Gelder, T. (1995). Mind as motion: Explorations in the dynamics of \ncognition . Cambridge, MA: MIT Press .", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "0b71a2c3-b090-4c63-9246-2c508d56a785": {"__data__": {"id_": "0b71a2c3-b090-4c63-9246-2c508d56a785", "embedding": null, "metadata": {"page_label": "99", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the relationship between adaptive dynamical systems and self-modeling network models in the context of mental processes?\n2. How can mathematical analysis be used to understand the dynamics and adaptivity of mental processes?\n3. What are some key concepts and theories related to the dynamics and adaptivity of mental processes that have been explored in the context of this document?", "prev_section_summary": "The section discusses the relationship between adaptive dynamical systems and self-modeling network models in the context of mental processes. Mathematical analysis can be used to understand the dynamics and adaptivity of mental processes, and examples of mathematical models that have been used to study this include the work of Funahashi and Nakamura, Kuipers and Kassirer, and Port and van Gelder. Other relevant works mentioned include W.R.'s \"Design for a Brain,\" Beer's \"Dynamical approaches to cognitive science,\" Descartes' \"The World,\" Kelso's \"Dynamic patterns,\" Kim's \"Philosophy of Mind,\" Kuipers' \"Commonsense reasoning about causality,\" Laplace's \"Philosophical Essays on Probabilities,\" Pearl's \"Causality,\" and Port and van Gelder's \"Mind as motion.\"", "section_summary": "The section discusses the relationship between adaptive dynamical systems and self-modeling network models in the context of mental processes. Mathematical analysis is used to understand the dynamics and adaptivity of mental processes. Key concepts and theories related to the dynamics and adaptivity of mental processes that have been explored in the context of this document include heterogeneous multi-modal graph neural networks, adaptive dynamical systems, self-modeling network models, and network-oriented modeling. The section also mentions several authors and their works related to these topics, including William Kaufman, Peter Laplace, Judea Pearl, Robert Port and Thomas van Gelder, K.R. Scherer, Eleanor Thelen and Linda Smith, Joris Treur, and his various works.", "excerpt_keywords": "1. Philosophy,\n2. Probabilities,\n3. Causality,\n4. Dynamics,\n5. Computational architecture,\n6. Emergent processes,\n7. Temporal factorization,\n8. Mediating state properties,\n9. Network modeling,\n10. Temporal-causal networks."}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "d64f1393-8ea3-4495-b60d-300797f050d1", "node_type": "4", "metadata": {"page_label": "99", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "e4a007ae349c076b957891f562c88e341d8c0e16c5ce8eaf008e24e30dd3c3cb"}, "2": {"node_id": "3ee06420-3e90-4e62-bdc6-a81fd84a1271", "node_type": "1", "metadata": {"page_label": "99", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a05214ed12213e7f306c2beacfea066665968eb8036cfc498d30eeabc43b70a9"}, "3": {"node_id": "99ac42b4-e3fc-40fc-9cc9-c28519b61054", "node_type": "1", "metadata": {"page_label": "99", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "1286b76484809c8cc50b4663f5e6ae3e35aa12c14c3f871b6fa63bb2babfcca9"}}, "hash": "fb85dd74236bf749c9d90b8acbf4fe018886e380c1fc8b28a9ddc1266e90eb0e", "text": "F.R.G. \n(William Kaufman, Los Altos, CA, 1983). \nLaplace, P.S. (1825) Philosophical Essays on Probabilities Springer-Verlag New York \n1995 Translated by AI Dale from the 5th French edition of 1825. \nPearl, J. (2000). 2000 . Causality: Cambridge University Press .  \nPort, R. F., & van Gelder, T. (1995). Mind as motion: Explorations in the dynamics of \ncognition . Cambridge, MA: MIT Press .  \nScherer, K. R. (2009). Emotions are emergent processes: They require a dynamic \ncomputational architecture. Phil. Trans. R. Soc. B, 364(1535), 3459 \u20133474 . \nThelen, E., & Smith, L. (1994). A dynamic systems approach to the development of cognition \nand action . Cambridge: MIT Press .  \nTreur, J. (2007a). Temporal Factorisation: A Unifying Principle for Dynamics of the \nWorld and of Mental States. Cognitive Systems Research, 8(2), 57\u201374. \nTreur J (2007b) Temporal Factorisation: Realisation of mediating state properties for \ndynamics Cognitive Systems Research 8 (2) (2007): 75-88. \nTreur, J. (2016a). Dynamic Modeling Based on a Temporal-Causal Network Modeling \nApproach. Biologically Inspired Cognitive Architectures, 16, 131\u2013168 (2016). \nTreur, J. (2016). Network-Oriented Modeling: Addressing Complexity of Cognitive, Affective \nand Social Interactions . Springer Publishers .  \nTreur, J. (2017). On the applicability of network-oriented modeling based on temporal- \ncausal networks: Why network models do not just model networks. J. Inf. \nTelecommun., 1, 23\u201340. \nFig.", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "99ac42b4-e3fc-40fc-9cc9-c28519b61054": {"__data__": {"id_": "99ac42b4-e3fc-40fc-9cc9-c28519b61054", "embedding": null, "metadata": {"page_label": "99", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the relationship between adaptive dynamical systems and self-modeling network models in cognitive modeling?\n2. How does network-oriented modeling based on temporal-causal networks address the complexity of cognitive, affective, and social interactions?\n3. What are the limitations of network models in modeling networks, and how can they be improved?", "prev_section_summary": "The section discusses the relationship between adaptive dynamical systems and self-modeling network models in the context of mental processes. Mathematical analysis is used to understand the dynamics and adaptivity of mental processes. Key concepts and theories related to the dynamics and adaptivity of mental processes that have been explored in the context of this document include heterogeneous multi-modal graph neural networks, adaptive dynamical systems, self-modeling network models, and network-oriented modeling. The section also mentions several authors and their works related to these topics, including William Kaufman, Peter Laplace, Judea Pearl, Robert Port and Thomas van Gelder, K.R. Scherer, Eleanor Thelen and Linda Smith, Joris Treur, and his various works.", "section_summary": "The section discusses the relationship between adaptive dynamical systems and self-modeling network models in cognitive modeling. It also explores the limitations of network models in modeling networks and how they can be improved. The section highlights the use of network-oriented modeling based on temporal-causal networks to address the complexity of cognitive, affective, and social interactions. The section also includes references to relevant research papers and the author's work on the topic.", "excerpt_keywords": "1. Network-oriented modeling\n2. Temporal-causal networks\n3. Cognitive architecture\n4. Affective interactions\n5. Social interactions\n6. Adaptive dynamical systems\n7. Self-modeling\n8. Utility\n9. Fairness\n10. Positivity"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "d64f1393-8ea3-4495-b60d-300797f050d1", "node_type": "4", "metadata": {"page_label": "99", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "e4a007ae349c076b957891f562c88e341d8c0e16c5ce8eaf008e24e30dd3c3cb"}, "2": {"node_id": "0b71a2c3-b090-4c63-9246-2c508d56a785", "node_type": "1", "metadata": {"page_label": "99", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "fb85dd74236bf749c9d90b8acbf4fe018886e380c1fc8b28a9ddc1266e90eb0e"}}, "hash": "1286b76484809c8cc50b4663f5e6ae3e35aa12c14c3f871b6fa63bb2babfcca9", "text": "Modeling Based on a Temporal-Causal Network Modeling \nApproach. Biologically Inspired Cognitive Architectures, 16, 131\u2013168 (2016). \nTreur, J. (2016). Network-Oriented Modeling: Addressing Complexity of Cognitive, Affective \nand Social Interactions . Springer Publishers .  \nTreur, J. (2017). On the applicability of network-oriented modeling based on temporal- \ncausal networks: Why network models do not just model networks. J. Inf. \nTelecommun., 1, 23\u201340. \nFig. 5.Role matrices specification for the self-modeling network model based on the example adaptive dynamical system.  J. Treur", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}, "c87cd0d2-0b34-4099-88fd-ef65e4c7708a": {"__data__": {"id_": "c87cd0d2-0b34-4099-88fd-ef65e4c7708a", "embedding": null, "metadata": {"page_label": "100", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the relationship between adaptive dynamical systems and self-modeling network models in the context of mental processes?\n2. How can mathematical analysis be used to model the dynamics and adaptivity of mental processes?\n3. What is the role of higher-order adaptivity in network-oriented modeling for adaptive biological, mental, and social network models?", "prev_section_summary": "The section discusses the relationship between adaptive dynamical systems and self-modeling network models in cognitive modeling. It also explores the limitations of network models in modeling networks and how they can be improved. The section highlights the use of network-oriented modeling based on temporal-causal networks to address the complexity of cognitive, affective, and social interactions. The section also includes references to relevant research papers and the author's work on the topic.", "section_summary": "The section discusses the relationship between adaptive dynamical systems and self-modeling network models in the context of mental processes. Mathematical analysis is used to model the dynamics and adaptivity of mental processes. The role of higher-order adaptivity in network-oriented modeling for adaptive biological, mental, and social network models is also discussed. The section references several papers and articles that explore the topic in more detail.", "excerpt_keywords": "1. Network science\n2. Adaptive networks\n3. Higher-order adaptivity\n4. Cognitive systems\n5. Temporal factorisation\n6. Criterial causation\n7. Informational content\n8. Free will\n9. Dynamical approach\n10. Cognitive science"}, "excluded_embed_metadata_keys": ["creation_date", "last_modified_date", "last_accessed_date"], "excluded_llm_metadata_keys": ["questions_this_excerpt_can_answer"], "relationships": {"1": {"node_id": "d5c086c9-cde5-47c6-b8d9-3bd77a6a1b8c", "node_type": "4", "metadata": {"page_label": "100", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30"}, "hash": "a4b8e0615edc8fa500aef85e9a3312b1119c7d2089b7fcdb874d64ed2cbfaaa4"}}, "hash": "512d26260e67554603c361325b9ab4472a691c8841518ac812d457fda66f7c7d", "text": "Cognitive Systems Research 70 (2021) 93\u2013100\n100Treur, J. (2020). Modeling Higher-Order Adaptivity of a Network by Multilevel Network \nReification. Network Science, 8(S1), S110\u2013S144. \nTreur, J. (2020). Network-oriented modeling for adaptive networks: Designing higher-order \nadaptive biological, mental and social network models. Springer Nature Publishers.  \nTreur, J. (2021). Modeling the Emergence of Informational Content by Adaptive \nNetworks for Temporal Factorisation and Criterial Causation. Cognitive Research \nJournal, 68, 34\u201352. Tse, P. U. (2013). The Neural Basis of Free Will: Criterial Causation. Cambridge: MIT Press.  \nvan Gelder, T. (1998). The dynamical hypothesis in cognitive science. Behavioral and \nBrain Sciences, 21(5), 615\u2013628. \nVan Gelder, T. and Port, R.F. (1995). It\u2019s about time: An overview of the dynamical \napproach to cognition. In: Port, R.F., & van Gelder, T. (1995). Mind as motion: \nExplorations in the dynamics of cognition. Cambridge, MA: MIT Press, pp. 1-43. J. Treur", "start_char_idx": null, "end_char_idx": null, "text_template": "[Excerpt from document]\n{metadata_str}\nExcerpt:\n-----\n{content}\n-----\n", "metadata_template": "{key}: {value}", "metadata_seperator": "\n"}, "__type__": "1"}}, "docstore/ref_doc_info": {"f40bb129-03c8-42b9-80ee-694879b4e4ae": {"node_ids": ["5ef454ee-9d20-4491-ba57-d0d4aa559569", "316f1e3e-adc4-4999-8696-fc25f0e20b50", "5c3e77ea-e060-4453-861a-af0a9c89a785", "cd77e6c2-e980-4dce-9375-a31aa31ccde4", "a11e5164-0951-41fc-9479-c9cccb09e5d8"], "metadata": {"page_label": "1", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the article \"Corporate Relative Valuation using Heterogeneous Multi-Modal Graph Neural Network\"?\n2. How does the availability of information about a company's invisible assets enable a new paradigm for learning and evaluating corporate relative values automatically?\n3. What is the methodology used in the article to develop a heterogeneous multi-modal graph neural network for corporate relative valuation?", "section_summary": "The section discusses the article \"Corporate Relative Valuation using Heterogeneous Multi-Modal Graph Neural Network\" by Yang et al. The purpose of the article is to develop a new paradigm for learning and evaluating corporate relative values automatically using the availability of information about a company's invisible assets, such as patents, talent, and investors. The methodology used in the article involves forming the companies and their core members as a heterogeneous graph with semantically-rich multi-modal data, and developing a heterogeneous multi-modal graph neural network method, named HM2, which deals with embedding challenges involving modal attribute encoding, multi-modal aggregation, and valuation prediction modules. The article highlights the effectiveness of the HM2 method in extracting latent embeddings that reflect domain experts' behavior and are effective for corporate relative valuation.", "excerpt_keywords": "1. Corporate relative valuation, 2. Heterogeneous graph neural network, 3. Multi-modal data, 4. Invisible assets, 5. Patents, talent, investors, 6. Domain experts, 7. Network embeddings, 8. Valuation prediction, 9. Non-publicly listed companies, 10. Venture capital firms."}}, "7d6b85a8-1c1c-4b4a-8c9f-9827ff561d18": {"node_ids": ["175214df-164e-4f15-baa4-95cce9ade065", "f33ded4d-b4d3-4685-93b3-bc7231ac85e8", "ab91e265-e1cc-463a-9277-8eb7f9086794", "045590d6-758e-4abb-86e8-8a3a02191a5f", "c83c8670-2419-4572-9a81-a859ba0021db"], "metadata": {"page_label": "2", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using a heterogeneous multi-modal graph neural network for corporate relative valuation?\n2. How does the use of cartoon characters in the example of company structure help protect privacy?\n3. What are the three factors that are considered in manual relative valuation without financial statements?", "prev_section_summary": "The section discusses the challenges in acquiring data for corporate relative valuation, especially for non-publicly listed companies. It also explains some of the methods used for corporate relative valuation and how professionals can use their expertise to carry out a comprehensive valuation of a company's core resources, members, and competitors. The section mentions the use of heterogeneous multi-modal graph neural networks for corporate relative valuation.", "section_summary": "The section discusses the use of a heterogeneous multi-modal graph neural network for corporate relative valuation (CRV). CRV is a method used to estimate the value of a company based on its structure and other relevant information. The section explains that CRV usually considers two aspects of a company's structure: affiliates and members. To protect privacy, the section uses cartoon characters to replace the companies and members. The section also discusses the limitations of manual CRV without financial statements and the need for automatic or semi-automatic CRV technology using machine learning models. The section mentions that manual CRV without financial statements considers three factors: the core resources of the company and its affiliates, the business conditions, and the competitive landscape.", "excerpt_keywords": "1. CRV, 2. company structure, 3. affiliate information, 4. member information, 5. cartoon characters, 6. heterogeneous graph, 7. statement analysis, 8. venture capital, 9. company valuation, 10. machine learning models."}}, "d543d099-ce2c-4224-b9a4-ad759f438763": {"node_ids": ["b3b9bab2-09dd-48d1-aede-2769a88f6bac", "61fa9f26-9135-48ed-9695-1adab3f9f588", "f300981f-69ac-4e0a-b4d7-aa69fe74db59", "b242dc15-7e7c-4e06-afa3-0a6a68e209c3", "c010ac8f-8b4b-44ba-8a15-52e55ecbdf4c", "d2f8d11b-631b-4434-8ae4-8da32d3a3437"], "metadata": {"page_label": "3", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of developing a deep heterogeneous graph method for Corporate Relative Valuation (CRV)?\n2. How does the heterogeneous multi-modal graph differ from a homogeneous graph in terms of node types and linkage representations?\n3. What are the different types of nodes and linkages in a heterogeneous multi-modal graph, and how are their attributes represented?", "prev_section_summary": "The section discusses the challenges associated with using traditional graph neural networks for corporate relative valuation analysis, the impact of heterogeneity in a multi-modal graph on the accuracy of the model, and the potential benefits of using a heterogeneous multi-modal graph neural network for corporate relative valuation analysis. The section also highlights the complexity and unstructured nature of the data in a heterogeneous multi-modal graph, and the limitations of directly concatenating company and member embeddings.", "section_summary": "The section discusses the development of a deep heterogeneous graph method for Corporate Relative Valuation (CRV) using a heterogeneous multi-modal graph. The heterogeneous multi-modal graph differs from a homogeneous graph in terms of node types and linkage representations, with different types of nodes having various attributes and linkages. The section also defines the heterogeneous multi-modal graph and provides an overview of the problem definition.", "excerpt_keywords": "heterogeneous graph, multi-modal information, content-associated graph, node set, linkage set, node type set, linkage type set, attributes, raw dimensional attribute representations, deep heterogeneous graph method, CRV."}}, "3290c02d-c27c-44d5-b841-aa705c70db9d": {"node_ids": ["13e77e0e-de3d-4c04-a7fd-9684826b4420", "82aa3f52-44ff-4ff9-8e85-01f0aa4f07dc", "f8bec8e6-d4da-4ede-96a8-c56cbb71f22f", "e43d10e5-2e85-4811-bca0-5cbb0dbf0907", "764a3f60-296f-4bbe-9f8d-611a57d7b14a"], "metadata": {"page_label": "4", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using a heterogeneous multi-modal graph neural network for corporate relative valuation?\n2. How does the graph representation of the dataset differ from a traditional graph representation, and what are the benefits of using node2vec embeddings for member nodes?\n3. What are the different types of linkages in the graph, and how are they represented?", "prev_section_summary": "The section discusses the challenges faced by companies in the internet industry when it comes to data collection and analysis, specifically in relation to corporate relative valuation. The heterogeneity of graphs in the internet industry differs from those in other domains, and this has implications for relative valuation. The potential benefits of using data from the internet industry for corporate relative valuation are discussed, including the fact that most emerging companies are in the internet industry, the internet industry takes up the largest number of companies in the collected data, and the heterogeneous graphs of internet companies have relevance island problem with the heterogeneous graphs of other domains. The section also notes that the data is mostly from startups or unlisted companies that are in need of relative valuation.", "section_summary": "The section discusses the use of a heterogeneous multi-modal graph neural network for corporate relative valuation. The graph representation of the dataset differs from a traditional graph representation, with node2vec embeddings used for member nodes. The graph consists of two types of nodes: company and core member, with three types of linkages: company-company, company-member, and member-member. The relative valuation level has four categories.", "excerpt_keywords": "HMMG, graph, nodes, attributes, company, core member, linkages, investment ratio, acquisition, one-hot encoding, member-member linkages, relative valuation level, categories, 100 millions, 200 millions, 300 millions, 400 million."}}, "f4ace53a-c604-4e78-a848-ab7cd073ebde": {"node_ids": ["80e29ef9-09f7-4ad0-b123-c0f084375603", "2f3fed68-fbc2-400e-bda1-721f056a8189", "43935afc-b148-4e05-a180-fc64c144d4d7", "00a8f7a7-a346-4dc4-b8aa-60cfeeeab5a2"], "metadata": {"page_label": "5", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the overall architecture of the Heterogeneous Multi-Modal Graph Neural Network (HM2) for corporate relative valuation?\n2. What are the different types of nodes and links in the HM2 model, and how are they encoded and aggregated?\n3. How does the HM2 model develop the loss for corporate valuation with structural triplet regularization?", "prev_section_summary": "The section discusses the challenges faced by Heterogeneous Multi-Modal Graph Neural Network (HM2) for Corporate Relative Valuation (CRV) task and presents HM2 as a solution. The three modules that make up HM2 are the Modal Attribute Encoding Module, Multi-Modal Aggregation Module, and Valuation Loss Module. The Modal Attribute Encoding Module encodes each type of neighbors separately after neighbor sampling, the Multi-Modal Aggregation Module adaptively aggregates different types of neighbors and input node itself, and the Valuation Loss Module considers the loss of valuation prediction and incorporates the similarities of different corporate node embeddings in the same business category.", "section_summary": "The section discusses the Heterogeneous Multi-Modal Graph Neural Network (HM2) for corporate relative valuation. The architecture of HM2 is described, including the different types of nodes and links, and how they are encoded and aggregated. The section also explains how the HM2 model develops the loss for corporate valuation with structural triplet regularization. The symbols used in the section are also defined.", "excerpt_keywords": "1. Heterogeneous Multi-Modal Networks\n2. Multi-Head Attention\n3. Adaptive Attention\n4. Structural Triplet Regularization\n5. Neighbor Sampling\n6. Modal Attribute Encoding\n7. Corporate Valuation\n8. Company-Company Linkage\n9. Company-Member Linkage\n10. Member-Member Linkage"}}, "5b21b249-9b0f-48d8-b500-41fb636aaa28": {"node_ids": ["f2712e0a-b6d7-462c-be38-6574eaf31e7f", "c61fc7bc-98e3-4656-aa93-f11f9245629b", "011cff06-fd54-4e89-a84c-14a498dfefc7", "2ebd71ad-f2b3-4aec-9f3b-505e92f90d1c", "826655f9-172c-46dc-99bf-1017a8305b63", "3b47d874-dd1a-42d2-a582-34bad3d8b0c9"], "metadata": {"page_label": "6", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Multi-Modal Graph Neural Network (HMMG) for Corporate Relative Valuation?\n2. How does the HMMG model handle the heterogeneity of nodes and their attributes in the graph?\n3. What are the benefits of using the HMMG model for corporate relative valuation compared to traditional GNN models?", "prev_section_summary": "The section discusses the use of Heterogeneous Multi-Modal Graph Neural Network (HM2) for corporate relative valuation. The purpose of using random walk sampling in HM2 for corporate relative valuation is to collect all types of neighbors and group them according to their frequency. The benefits of using HM2 for corporate relative valuation compared to traditional methods that only sample direct neighbors include the ability to aggregate heterogeneous information that contains different modal features and the ability to handle various feature types and dimensions. The section describes the random walk sampling process in detail, which involves sampling a fixed size of neighbors with random walk sampling and grouping different types of neighbors according to their frequency.", "section_summary": "The section discusses the use of a Heterogeneous Multi-Modal Graph Neural Network (HMMG) for corporate relative valuation. The purpose of the HMMG model is to handle the heterogeneity of nodes and their attributes in the graph, and to collect high-order member neighbors for each node when constructing neighbors. The model focuses on the embedding learning of corporate nodes, but also takes into account the impact of different types of neighbors on node embeddings. The section also mentions the benefits of using the HMMG model for corporate relative valuation compared to traditional GNN models.", "excerpt_keywords": "1. Graph Neural Networks (GNNs)\n2. Heterogeneous Graphs\n3. Corporate Nodes\n4. Member-Member Linkage\n5. Neighbor Embedding\n6. Modal Attribute Encoding\n7. Multi-Modal Neighbors\n8. Core Resources\n9. Valuation\n10. Deep Graph Neural Networks"}}, "b26ce5b7-5cd1-4c4f-ab00-7f0ec4074897": {"node_ids": ["edd73257-0ab1-45ab-abc8-54d97a7edd0a", "d8e60937-b046-4729-83f5-35b2bb6e3498", "e40e0d0c-2842-4989-9e31-f6cbe20c35ab", "4bb4fea5-d8cc-4249-99cb-d5a06b1528ac", "60bde3cc-3c74-444d-808f-5d7a1bbcea2d", "f961527e-ce85-43f9-a7fc-dbf858464d20"], "metadata": {"page_label": "7", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using adaptive weights in the heterogeneous multi-modal graph neural network for corporate relative valuation?\n2. How is the heterogeneous graph representation structural loss incorporated into the overall loss function for training the HM2 model?\n3. What is the role of the mapping introduced in Section 3.1 in the heterogeneous multi-modal graph neural network for corporate relative valuation?", "prev_section_summary": "The section discusses the use of a Heterogeneous Multi-Modal Graph Neural Network (HMMGNN) for Corporate Relative Valuation. The purpose of the HMMGNN is to improve the final representation of input nodes by aggregating different modal embeddings using a Multi-Modal Aggregating Module. The Multi-Modal Aggregating Module uses adaptive attention to capture more discriminative feature capability and effectively overcome two problems mentioned in the section. The two problems that the learnable weight \u03b1 in the HMMGNN helps to overcome are the impact of direct and non-direct neighbors, and the ability to measure relationships between input nodes and different types of neighbors. The section also provides equations for the embedding output of t\u2212th type of neighboring nodes and the final representation of v1i.", "section_summary": "The section discusses the use of a heterogeneous multi-modal graph neural network (HM2) for corporate relative valuation. The network uses adaptive weights to discover relationships among different modalities, and a heterogeneous graph representation structural loss is incorporated into the overall loss function for training. The mapping introduced in Section 3.1 is also discussed, which is used to concatenate embeddings from different modalities. The section concludes with a multi-task learning approach that combines corporate relative valuation loss and heterogeneous graph representation structural loss for training the HM2 model.", "excerpt_keywords": "1. Adaptive weights, 2. Modal embeddings, 3. Leaky ReLU, 4. Rectified Linear Unit, 5. Parameter, 6. Concatenated embeddings, 7. Heterogeneous graph representation, 8. Multi-task learning, 9. Discriminative representation, 10. Overall loss function."}}, "8500b2d6-015e-436e-a558-aa4efc073e92": {"node_ids": ["035a0bb6-ce4b-4b04-bf6e-bda7e200b914", "eda5aba7-a3f1-4bf9-9dd3-eaa15059a4aa", "ab216420-b362-48ea-8290-5b42c8350509", "e9a91b17-9249-41b8-8ec9-c8c477194fcb", "c5b7cc46-849c-465a-8715-8d69c0f9334d", "bd4d0ef9-9a64-4eca-9203-80315058db15", "f8b751bb-40f9-4123-b6fc-5575dc5f2383", "c7ae3fa8-e717-4a52-9b40-c15d32848a2b"], "metadata": {"page_label": "8", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the accuracy of the Heterogeneous Multi-Modal Graph Neural Network (HetGNN) in predicting corporate relative valuation for different training data ratios?\n2. How does the performance of the HetGNN compare to other machine learning algorithms such as Support Vector Machines (SVM), Multi-Layer Perceptron (MLP), and K-Nearest Neighbors (KNN) in predicting corporate relative valuation?\n3. What is the precision of the HetGNN in predicting corporate relative valuation for different training data ratios?", "prev_section_summary": "The section discusses a heterogeneous graph neural network model for corporate relative valuation. The model constructs two modules to aggregate feature information of heterogeneous nodes, with the first module learning embeddings of heterogeneous contents using an LSTM module, and the second module aggregating embeddings of different neighboring types. The section also compares the heterogeneous graph neural network model to the m2vec model in terms of their approach to generating heterogeneous neighborhoods.", "section_summary": "The section discusses the accuracy and performance of a Heterogeneous Multi-Modal Graph Neural Network (HetGNN) in predicting corporate relative valuation. The study compares the performance of HetGNN to other machine learning algorithms such as Support Vector Machines (SVM), Multi-Layer Perceptron (MLP), and K-Nearest Neighbors (KNN) in predicting corporate relative valuation. The study also examines the precision of HetGNN in predicting corporate relative valuation for different training data ratios. The results are presented in Table 3, which shows the accuracy and precision of each algorithm for different training data ratios.", "excerpt_keywords": "1. Corporate relative valuation prediction\n2. SVM\n3. MLP\n4. KNN\n5. HetGNN\n6. GAT\n7. SAGE\n8. ASNE\n9. m2vec\n10. HAN"}}, "c69dbb5a-e5cd-4e5b-baec-971425360fe2": {"node_ids": ["ee7c1408-a55f-44ee-ac96-0fb40bbda6b8", "7f8e09db-ab4a-4a4e-8ecf-b23c92a915c9", "dbaf40b4-9063-4dd0-8a24-da12f3dc98cd", "faf2efec-9e29-4db3-8c8a-ec87e4f3548e", "4b8db73f-108b-441c-be29-3db8f6bc5d78"], "metadata": {"page_label": "9", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the article \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\"?\n2. How does the proposed method, HM2, compare to traditional linear methods in terms of performance for corporate relative valuation?\n3. What are the evaluation metrics used to assess the performance of the proposed method, HM2, in the corporate relative valuation task?", "prev_section_summary": "The section discusses the use of a heterogeneous multi-modal graph neural network for corporate relative valuation. The purpose of this approach is to improve the accuracy of valuation by incorporating multiple types of data, such as financial and non-financial information. The validation set is randomly selected from the training set for cross-validation, and the \u03bbin Equation 7 is used to tune the hyper-parameters. The section also outlines the methodology for comparing different methods and the stopping criteria for training.", "section_summary": "The section discusses a research article titled \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\" by Yang et al. The purpose of the article is to propose a method, HM2, for corporate relative valuation using a heterogeneous multi-modal graph neural network. The proposed method is compared to traditional linear methods in terms of performance, and evaluation metrics such as Accuracy, Recall, Precision, and F1-measure are used to assess the performance of HM2. The results show that graph embedding methods outperform traditional linear methods and most methods achieve good performance in the corporate relative valuation task.", "excerpt_keywords": "1. HM2,\n2. Graph embedding,\n3. Corporate relative valuation,\n4. Node classification,\n5. Accuracy,\n6. Recall,\n7. Precision,\n8. F1-measure,\n9. Data imbalance,\n10. Nvidia 2080ti."}}, "721e6fe2-2be3-4029-8b16-c9648c3e2621": {"node_ids": ["ba8f4e68-acd7-4a47-a776-80b310febb8f", "ef19d03d-3442-458c-adef-cb8a162d42bb", "14a82aaf-27b8-4cf4-b149-7e5d5eceb628", "228042aa-b36d-4bfe-a3a7-e4c838cb1a8d", "33d77f1b-639c-42f9-8483-4026f5c09bef", "f5231109-65d5-43e0-8568-e77d8249a5ee", "4628dd81-f2b7-4da3-8457-109259410331", "caf9301f-cc5a-4f40-bfae-626ad1968ad7"], "metadata": {"page_label": "10", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the performance of the Heterogeneous Multi-Modal Graph Neural Network (HM2-N) model in predicting corporate relative valuation with different training data ratios?\n2. How does the precision of the HM2-N model vary with different training data ratios?\n3. What is the impact of the heterogeneity of the multi-modal graph on the performance of the HM2-N model in predicting corporate relative valuation?", "prev_section_summary": "The section discusses a study that aims to solve the problem of corporate relative valuation using a Heterogeneous Multi-Modal Graph Neural Network (HM2). The study compares the performance of different variants of HM2, including HM2-R, HM2-A, and HM2-B, to other methods for corporate relative valuation. The results reveal that HM2 outperforms other methods, demonstrating the effectiveness of neighbor sampling and linkage-aware multi-head attention based encoding. The study is reported in tables 5 and 6, and the variant of HM2 that replaces the trans-former based attribute encoding module with a fully connected network to encode heterogeneous nodes is particularly effective.", "section_summary": "The section discusses the performance of a Heterogeneous Multi-Modal Graph Neural Network (HM2-N) model in predicting corporate relative valuation with different training data ratios. The study compares the accuracy and precision of the HM2-N model with other models (HM2-L, HM2-FC, HM2-A, HM2-B, HM2-R, and HM2) using an ablation study. The results show that the HM2-N model outperforms the other models in terms of accuracy and precision, with the best results achieved when using a training data ratio of 70%. The heterogeneity of the multi-modal graph has a significant impact on the performance of the HM2-N model, with better results achieved when the graph is more heterogeneous.", "excerpt_keywords": "1. Ablation study\n2. HM2-N\n3. HM2-L\n4. HM2-FC\n5. HM2-A\n6. HM2-B\n7. HM2-R\n8. HM2\n9. Accuracy\n10. Precision"}}, "0c16360d-0190-4875-9f56-c31cfa1645d9": {"node_ids": ["77d4247d-6f97-4b5e-af12-9949a43daab5", "1a36b08b-4d79-4031-8b3e-cb2cd47bf4e2", "d4cc0b60-4559-4a49-b3e2-48aa039b2322", "837a6005-5f03-4a63-a9d1-4e2092bbe90b", "de05702f-74c6-4c8d-a3d6-2c6af282b163"], "metadata": {"page_label": "11", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the study in the article \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\"?\n2. How does the size of the sampled neighbors set affect the performance of the Heterogeneous Multi-Modal Graph Neural Network (HM2) in the study?\n3. What is the impact of the dimension of the embeddings on the performance of the Heterogeneous Multi-Modal Graph Neural Network (HM2) in the study?", "prev_section_summary": "The section discusses the use of heterogeneous multi-modal graph neural networks for corporate relative valuation. The significance of member-member linkage in this context is explored, as well as the role of core members in the company valuation process. The study finds that the effect of member-member linkage is weak, and that only direct members are necessary for company valuation. The performance degradation of removing member-member linkage is more significant than removing core members.", "section_summary": "The section discusses a study on using a Heterogeneous Multi-Modal Graph Neural Network (HM2) for corporate relative valuation. The purpose of the study is to learn more discriminative embeddings of company nodes. The size of the sampled neighbors set and the dimension of the embeddings are analyzed in the study. The study shows that the best neighbor size is between 10 and 15, and the best embedding dimension is between 32 and 256. The study also shows that HM2 is superior to without embedding, indicating that it is more meaningful to consider the type of linkages when learning node embedding.", "excerpt_keywords": "1. Node embedding, 2. Hyper-parameter study, 3. Evaluation metrics, 4. Classification task, 5. Regression task, 6. Sampled neighbors set, 7. Neighbor size, 8. Dimension, 9. Discriminative embedding, 10. Linkage type"}}, "9e4a0704-7ac7-4488-a38b-21ebb11fe736": {"node_ids": ["384fb0aa-5011-4a5f-a17c-4e27b0d2d10e", "76dfbd6e-2f7c-4dad-ba67-b8f0541ebd08", "9a79589e-b750-4e08-9e6c-0d47148f7d7c", "5d153fee-4bcd-46eb-be08-e668159ce954"], "metadata": {"page_label": "12", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using a heterogeneous multi-modal graph neural network for corporate relative valuation?\n2. How does the multi-modal aggregation module and the modal attribute encoding module contribute to the reliability of the prediction in the heterogeneous multi-modal graph neural network for corporate relative valuation?\n3. What are the benefits of using attention visualizations in the heterogeneous multi-modal graph neural network for corporate relative valuation?", "prev_section_summary": "The section discusses the challenges associated with obtaining full financial statements and stock information for corporate relative valuation, and how analyzing a company's core resources and members can be used as an alternative method. It also introduces multi-modal learning and its approach to improving performance in corporate relative valuation. The section discusses the different approaches to multi-modal fusion and their assumptions, and mentions related techniques for valuation.", "section_summary": "The section discusses the use of a heterogeneous multi-modal graph neural network for corporate relative valuation. The purpose of using this network is to improve the reliability of the prediction by incorporating multiple types of data and relationships. The multi-modal aggregation module and the modal attribute encoding module are key components of the network that contribute to its effectiveness. Attention visualizations are also discussed as a way to better understand the network's predictions. The section includes examples of attention visualizations for two cases.", "excerpt_keywords": "1. Multi-modal aggregation module\n2. Modal attribute encoding module\n3. Attention visualization\n4. Reliable prediction\n5. Personalized recommendation\n6. Neighborhood-based recommendation\n7. Company neighbors\n8. Member neighbors\n9. Fairness and positivity\n10. Utmost utility"}}, "8f18d25c-ff85-49d7-a478-11d026ab472f": {"node_ids": ["b163aa80-b9e0-45d9-ac32-f3d3df8ff57e", "d50f733c-49de-4759-8e76-2362aa141eca", "19022137-33ca-4bc4-9c25-84583e635017", "1cbbd6e3-e7ce-47d5-a560-d15fca7ecc58", "c81e1197-5c10-4242-9505-acf800451254", "6283fc3c-7275-4b4e-8d9b-50df3aebb288", "57e041d4-db81-42b9-b181-900c2666dba8", "f5ded79c-9fee-4f07-8065-576f28712bea", "062d9cab-f298-462f-af4b-0c2cfb52a456"], "metadata": {"page_label": "13", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the article \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\"?\n2. What is the significance of heterogeneous multi-modal graph neural networks in the field of corporate relative valuation?\n3. How does the proposed method in the article differ from existing methods for corporate relative valuation?", "prev_section_summary": "The section discusses the development of HM2, an HGNN style method, for corporate relative valuation (CRV). HM2 differs from multi-instance based methods for CRV in terms of considering relationships among nodes. The triplet loss with embedding of competitors is used in HM2's learning process to learn more discriminative features. The effectiveness of HM2 is demonstrated through extensive experiments on real-world CRV data. The research was supported by various funding sources.", "section_summary": "The section discusses the article \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\" by Yang et al. The purpose of the article is to propose a method for corporate relative valuation using a heterogeneous multi-modal graph neural network. The significance of heterogeneous multi-modal graph neural networks in the field of corporate relative valuation is highlighted, as they allow for the integration of multiple types of data and the capture of complex relationships between entities. The proposed method differs from existing methods for corporate relative valuation in that it uses a graph neural network to model the relationships between entities and incorporates multiple types of data.", "excerpt_keywords": "1. Personalized question recommender system,\n2. Intelligent job interview,\n3. Talent management,\n4. Structure-aware convolutional neural network,\n5. Deep learning,\n6. Airbnb search,\n7. Agency costs,\n8. Corporate finance,\n9. Takeovers,\n10. Free cash flow."}}, "b1879836-b599-41e1-af94-4bfe72a77860": {"node_ids": ["39e4ff77-a4f5-498e-a2fb-94f577ae7908", "a4b15fbe-d158-4265-b5ef-7acf0309dbb3", "856e13fe-3def-4ed6-b440-d474562d6790", "3f8b5798-e6c2-4a6e-b2f0-a8c043396824", "8cfb97d7-4310-4249-9ead-f100246bb15f", "7d0469e6-407b-40f2-8c73-5d72911d8f27"], "metadata": {"page_label": "14", "file_name": "Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "file_path": "docs\\Corporate Relative Valuation Using Heterogeneous MultiModal Graph Neural Network Yang et al.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the title of the article and who are the authors?\n2. What is the journal that the article has been accepted for publication in?\n3. What is the research focus of the authors and what are some of their previous publications in this area?", "prev_section_summary": "The section discusses the use of heterogeneous multi-modal graph neural networks for corporate relative valuation. The key topics include the difference between heterogeneous multi-modal graph neural networks and traditional graph neural networks, the role of meta-path based context in top-N recommendation with a neural co-attention model, and the use of semi-supervised multi-modal learning for corporate relative valuation. The entities mentioned in the section include Y. Yang, K. Wang, D. Zhan, H. Xiong, and Y. Jiang, P. Cui, X. Wang, J. Pei, and W. Zhu, Y. Sun, J. Han, C. C. Aggarwal, and N. V. Chawla, B. Hu, C. Shi, W. X. Zhao, and P. S. Yu, and IEEE Xplore.", "section_summary": "The section discusses a research article titled \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\" by Yang et al. The article has been accepted for publication in the journal Transactions on Knowledge and Data Engineering. The authors' research interests include machine learning and data mining, specifically heterogeneous learning, model reuse, and incremental mining. The authors have published over 10 papers in leading international journals and conferences, and serve as PCs in conferences such as IJCAI, AAAI, ICML, and NIPS. The article focuses on the use of heterogeneous multi-modal graph neural networks for corporate relative valuation.", "excerpt_keywords": "1. Heterogeneous network embedding,\n2. Attributed multiplex heterogeneous network embedding,\n3. Multi-source information fusion,\n4. Machine learning,\n5. Data mining,\n6. Incremental mining,\n7. Model reuse,\n8. Novel software technology,\n9. National Key Lab,\n10. Nanjing University."}}, "e4af41c9-853f-4975-8439-415e32a44762": {"node_ids": ["ab357d46-6c15-4fb7-98af-1eec2922bd7f", "d268d7b3-399c-43ba-b818-05417e289f3c"], "metadata": {"page_label": "474", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using enterprise architecture (EA) models for decision making in planning purposes?\n2. How does a gap analysis help in identifying areas that need improvement in an enterprise architecture?\n3. What is the role of a transformation model and an action repository in creating transformation paths towards a desired and detailed target architecture?", "prev_section_summary": "The section discusses Dr. Yi Yang's current research interests and affiliations, as well as his scientific publications and citations. Dr. Yang is a Chang-Jiang Professor with the School of Computer Science and Engineering at NUST in Newark, NJ, USA. His research interests include pattern recognition, computer vision, and machine learning. He has authored over 200 scientific papers and has been cited more than 6000 times in the Web of Science and 15,000 times in Scholar Google. Dr. Yang is a Fellow of IAPR and serves as an Associate Editor of several journals and conferences, including Pattern Recognition, Pattern Recognition Letters, the IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS, and Neurocomputing.", "section_summary": "The section discusses the use of enterprise architecture (EA) models for decision making in planning purposes. It explains how a gap analysis can help identify areas that need improvement in an enterprise architecture. The role of a transformation model and an action repository in creating transformation paths towards a desired and detailed target architecture is also described. The paper presents a use case example and proposes a technical realization of the solution. Key topics and entities discussed include enterprise architecture management, enterprise architecture, gap analysis, transformation model, action repository, and graph transformation.", "excerpt_keywords": "Enterprise architecture planning, gap analysis, transformation model, graph transformation, Enterprise architecture management, EA models, business, processes, integration, software, technology, managed approach, complexity, abstraction, redesign, enterprise, planning, decision making, change management, adaptation, evolution, organization, components, relationships, environment, principles, guided design, controlled redesign, complexity management."}}, "3c2e8b58-e9eb-4f7a-884f-41f0155d56ef": {"node_ids": ["81acb7c7-dcff-4da1-b3d0-259f2d86fa79", "fc8366b8-f6a1-4bd1-ade4-041f326ac4dc"], "metadata": {"page_label": "475", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What are the different decision levels involved in EA planning activities and how do they vary in detail and levels of abstraction?\n2. How can gaps be derived from two EA models for different points in time, and what is the transformation model by Aier and Gleichauf used to connect architectural building blocks from these models?\n3. What is the proposed technical realization for the solution to get from gaps to transformation paths based on semantic web technologies and graph transformations?", "prev_section_summary": "The section discusses the key components of an Enterprise Architecture (EA) model and their role in supporting decision making for planning purposes. It also explores the complexity of an EA and the challenges it presents for enterprises in terms of steering and controlling its redesign. The section highlights the importance of a managed approach in EA planning and how it can help enterprises cope with the complexity of their systems. The section also discusses the level of abstraction, the number of stakeholders involved, and the change of internal and external conditions inherent to EAs as factors contributing to the complexity of an EA.", "section_summary": "The section discusses the different decision levels involved in Enterprise Architecture (EA) planning activities and how they vary in detail and levels of abstraction. It also describes how gaps can be derived from two EA models for different points in time and introduces a transformation model by Aier and Gleichauf to connect architectural building blocks from these models. The section proposes a technical realization based on semantic web technologies and graph transformations to get from gaps to transformation paths. The foundations of EA models and their usage for planning purposes are also introduced, along with semantic web technologies and graph transformations for planning purposes.", "excerpt_keywords": "Enterprise Architecture, Planning, Models, Gaps, Transformation, Architectural Building Blocks, Action Repository, Semantic Web Technologies, Graph Transformations, Management Cycle."}}, "9d4c0667-0535-499e-afcc-b195f78cb549": {"node_ids": ["813b7a2f-e1b4-4784-96f3-1be0370c1b22", "f00a11b4-bbda-48c0-a484-294f58a4e657"], "metadata": {"page_label": "476", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the difference between a gap analysis and a delta analysis in enterprise architecture planning?\n2. How do semantic web technologies help integrate heterogeneous data sets and formalize the underlying structure of information?\n3. What are the two standards relevant for a proposed technical realization using semantic web technologies?", "prev_section_summary": "The section discusses the management cycle for Enterprise Architecture Models (EAM) and how models can support the plan phase of this cycle. EA models can be used to describe an enterprise's architecture at different points in time and guide the development of an EA from the current towards a target architecture. The development of a target architecture is influenced by business requirements, strategic goals, and IT objectives such as master data consolidation, improving IT flexibility, and driving the coverage of standard platforms. The section also introduces semantic web technologies and graph transformations for planning purposes.", "section_summary": "The section discusses the difference between a gap analysis and a delta analysis in enterprise architecture planning, and how semantic web technologies help integrate heterogeneous data sets and formalize the underlying structure of information. The two standards relevant for a proposed technical realization using semantic web technologies are the Web Ontology Language (OWL) and the SPARQL Query Language for RDF. The Resource Description Framework (RDF) is a basis for both standards, and an RDF graph consists of triples of the form 'subject, predicate, object'.", "excerpt_keywords": "1. Architecture method, 2. Gap analysis, 3. Semantic web technologies, 4. Ontology, 5. Web Ontology Language (OWL), 6. SPARQL Query Language, 7. Resource Description Framework (RDF), 8. Information querying, 9. Machine understanding, 10. Heterogeneous data sets."}}, "fc130a71-3273-495a-8425-cb428aafe536": {"node_ids": ["3b03f2dd-94df-43a3-9f25-751fd859ee56", "560eaa93-8cb6-4f32-b481-beb417607cf5"], "metadata": {"page_label": "477", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using graph transformations in enterprise architecture planning?\n2. How does the proposed approach of modeling current and target architecture help in generating a transformation path?\n3. What is the role of business support maps in ensuring the comparability of the models in the proposed approach?", "prev_section_summary": "The section discusses the use of graph neural networks in planning approaches for Enterprise Architecture (EAM). It explains the difference between an RDF graph and a task network in planning approaches for EAM, and how semantic web technologies have been applied to domains of interest, including EAM. The section also discusses state space model based planning approaches used in EAM and how they differ from task networks. Finally, the section mentions some existing implementations of semantic web technologies for EAM, including TopQuadrant's TopBraid Composer and Essential Project.", "section_summary": "The section discusses the use of graph transformations in enterprise architecture planning. Graph transformations are used to solve planning problems by applying graph transformations to a model until a solution is found. The proposed approach involves modeling the current and target architecture at the same level of detail to ensure comparability. Business support maps are used to relate applications to supported processes and organization units. The result of the modeling phase are two sets: currentArchitecture and targetArchitecture.", "excerpt_keywords": "1. Graph transformations\n2. Planning problem\n3. State space\n4. Computation time\n5. Graph patterns\n6. Target architecture\n7. Domain expert\n8. Modeling\n9. Business support maps\n10. EA model"}}, "9bc994d8-83ee-4424-aa7d-cab1758f2d15": {"node_ids": ["48260b1b-0415-4f80-b0e2-30144e171771", "f8d21845-0325-4ffa-a0b5-0718e6079c93"], "metadata": {"page_label": "478", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of performing a gap analysis in architecture building blocks?\n2. How does the successor relationship model work in a transformation model for architecture building blocks?\n3. What are the three subsets identified during a gap analysis in architecture building blocks?", "prev_section_summary": "The section discusses the purpose of business support maps in restoring the comparability of models in enterprise architecture planning. It also explains the core of an enterprise architecture model consisting of three types of elements: architecture building blocks, relations between architecture building blocks, and attributes of architecture building blocks. The section also explains the difference between the current architecture and target architecture in enterprise architecture planning and how they are modeled in the solution.", "section_summary": "The section discusses the purpose of performing a gap analysis in architecture building blocks, how the successor relationship model works in a transformation model for architecture building blocks, and the three subsets identified during a gap analysis in architecture building blocks. The section also explains the definition of the transformation model and the successor relationships.", "excerpt_keywords": "architecture, building blocks, relations, attributes, gap analysis, current architecture, target architecture, stable, transformation model, successor relationships."}}, "bd8e03bc-e601-4841-824f-ed9aeec49953": {"node_ids": ["7e4e6d40-6aa8-4450-845d-c01cb3d0879c", "8ee342f5-c56f-40b6-8ab3-374f48f95ecc"], "metadata": {"page_label": "479", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the noSuccessor and noPredecessor sets in the transformation model of the current architecture?\n2. How does the model of the current architecture consider applications, services, and business building blocks to generate suggestions for a detailed target architecture?\n3. What is the process for overruling inappropriate suggestions made by the model of the detailed target architecture for a domain expert?", "prev_section_summary": "The section discusses the transformation model in the context of building block successor relationships in enterprise architecture planning. The transformation model is defined as the set of building blocks, relations, and attributes that are common to the current and target architectures. Successor relationships are used to identify the successor type for building blocks in the current and target architectures. The different types of successor relationships that can be identified in the transformation model are noSuccessor, noPredecessor, oneToOne, oneToMany, manyToOne, and manyToMany. The inverse of the successor relation is the predecessor relation. The section also mentions the importance of identifying the successor type for building blocks in order to plan for the transformation of the architecture.", "section_summary": "The section discusses the purpose of the noSuccessor and noPredecessor sets in the transformation model of a current architecture, how the model considers applications, services, and business building blocks to generate suggestions for a detailed target architecture, and the process for overruling inappropriate suggestions made by the model for a domain expert. The section also explains the different subsets of building blocks in the transformation model and how the model generates suggestions for a detailed target architecture by following a suggestion and processing all sets of successor relationships.", "excerpt_keywords": "1. noSuccessor set\n2. provided services\n3. target architecture\n4. applications\n5. dependencies\n6. detailed target architecture\n7. successor relationships\n8. business building blocks\n9. implementation independent\n10. domain expert"}}, "4d5cdb46-0352-41e4-9e9c-957f124e0967": {"node_ids": ["74e7fd9b-b060-48cf-80ce-3e8c7bd02324", "b1239ece-1757-4579-9a39-e084b8c6f1c8"], "metadata": {"page_label": "480", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the \"oneToMany\" set in the context of enterprise architecture planning?\n2. How does the \"manyToOne\" set suggest that successors provide services in the target architecture?\n3. What is the role of the domain expert in the enterprise architecture planning process?", "prev_section_summary": "The section discusses the use of a Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation in generating suggestions for a target architecture. The model of a detailed target architecture is generated by processing all sets of successor relationships and allowing domain experts to overrule inappropriate suggestions. The suggestions for provided services are generated by checking if they are used by an application that is part of the target architecture or the consuming application has a successor relationship. If there are no successor or predecessor relationships, a manual addition of provided services and their business building blocks in the target architecture is necessary.", "section_summary": "The section discusses the role of the \"oneToMany\" and \"manyToOne\" sets in enterprise architecture planning, as well as the role of the domain expert in the process. The \"oneToMany\" set suggests that successors provide services in the target architecture, while the \"manyToOne\" set suggests that the successor provides services already provided in the current architecture and provides all services of the other predecessors. The domain expert is responsible for modeling additional services and ensuring that all provided services have been modeled in the target architecture.", "excerpt_keywords": "1. Architecture,\n2. Succession,\n3. Transformation,\n4. Business building blocks,\n5. Services,\n6. Predecessor,\n7. Successor,\n8. OneToMany,\n9. ManyToOne,\n10. ManyToMany."}}, "2e1c4628-7cdd-4f08-8d86-072b77ae9761": {"node_ids": ["ae7bdbae-9d01-44e6-8f71-00f59971ea39", "0d9a1aed-8f69-4212-80e7-44708a37f5e7"], "metadata": {"page_label": "481", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the \"oneToMany set\" in the guided refinement process for enterprise architecture planning?\n2. How does the \"noPredecessor set\" and \"noSuccessor set\" affect the suggested services for an application in the target architecture?\n3. What is the role of the domain expert in the guided refinement process for enterprise architecture planning?", "prev_section_summary": "The section discusses the purpose of modeling additional services in the target architecture and how it contributes to sound transformation planning. It also explains the difference between the oneToOne and manyToOne sets and their role in suggesting used services to be used in the target architecture. The section also highlights the domain expert's role in suggesting used services and the importance of storing information about successor relationships of the services. Additionally, the section discusses the use of manyToMany, oneToOne, and manyToOne sets to suggest used services and the ability for the domain expert to choose which services are used by multiple successors.", "section_summary": "The section discusses the guided refinement process for enterprise architecture planning, specifically focusing on the \"oneToMany set,\" \"noPredecessor set,\" and \"noSuccessor set.\" The purpose of these sets is to suggest services for an application in the target architecture based on the current architecture. The role of the domain expert in this process is to model additional used services for each application. The results of the guided refinement process include a detailed target architecture with provided and used services, and related business building blocks. Gap analysis can be performed again to identify any remaining gaps between the current and target architecture. An action repository is also discussed, which is used to describe possible changes in the transformation path from the current to the target architecture.", "excerpt_keywords": "1. Enterprise Architecture Planning\n2. OneToMany set\n3. Predecessor\n4. Successor\n5. Gaps\n6. Transformation Paths\n7. Action Repository\n8. Abstract actions\n9. Preconditions\n10. Sequencing of actions"}}, "6de21fe1-d7d1-47c2-be30-c83697669eb5": {"node_ids": ["f64a2da0-63e9-404d-9e95-e8d462c4da03", "5877d42f-f6c7-4b0c-a003-cc67dd788599"], "metadata": {"page_label": "482", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the difference between atomic and composed actions in the context of enterprise architecture planning?\n2. How does the logical order of abstract actions prevent the creation of loops in the transformation path?\n3. Can you explain how the action repository and transformation model are used to create possible transformation paths in enterprise architecture planning?", "prev_section_summary": "The section discusses the purpose and role of an action repository in the Enterprise Architecture (EA) model. The action repository is used to model abstract actions that specify preconditions and effects for changes to an architecture model. The actions are described on an abstract level and can be checked for impact if the meta-model of the EA changes. The section also explains how the abstract action matches with the concrete model of the different states through a graph pattern.", "section_summary": "The section discusses the concepts of atomic and composed actions in the context of enterprise architecture planning, and how the logical order of abstract actions prevents the creation of loops in the transformation path. It also explains how the action repository and transformation model are used to create possible transformation paths in enterprise architecture planning. The section highlights the importance of modeling abstract actions for shutting down and developing building blocks and abstract actions that take care of the relationships between the building blocks and the attributes of the building blocks. The logical order of abstract actions prevents the creation of loops in the transformation path and ensures that the predecessor services are not shut down until the successor service is developed.", "excerpt_keywords": "1. Architecture model\n2. Abstract actions\n3. Concrete actions\n4. Transformation path\n5. Planning knowledge base\n6. Logical order\n7. Dependencies\n8. Service\n9. Application\n10. Development"}}, "0de2524a-66fb-45c2-8a35-7ad0f2e66e36": {"node_ids": ["5eb87bf6-d65a-494f-9093-4eb198babbc3", "6672141f-5ff2-44e3-95d9-49dd3bfd4b97"], "metadata": {"page_label": "483", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation?\n2. How does the selection process for choosing concrete actions in the transformation path enhancement work?\n3. What is the role of master data management in the development of applications in an organization's IT landscape?", "prev_section_summary": "The section discusses the use of a breadth search versus a depth search in creating transformation paths in enterprise architecture planning. It also explains how the action repository and the transformation model aid in the creation of possible transformation paths. The logical order for developing a new service in enterprise architecture planning is also discussed, as well as the process of creating the transformation path using the action repository, the transformation model, and the models of the current and target architectures.", "section_summary": "The section discusses the use of a Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation in enterprise architecture planning. The network is used to determine all possible transformation paths from the current to the target architecture and to select concrete actions for the transformation path enhancement process. The section also discusses the role of master data management in the development of applications in an organization's IT landscape and provides a use case for the introduction of master data management in the research and development division of an organization. The use case shows a part of the model of the current architecture of the organization's IT landscape and the development of a master data management system to provide services to other applications.", "excerpt_keywords": "Enterprise architecture planning, state space, transformation paths, concrete actions, development costs, maintenance costs, benefits, risks, resource constraints, master data management, development master data management, research and development division, IT landscape, applications, data consistency, redundancy, outdated data, organization, model, current architecture, development master data management system, services, DevManager, product planning tool, quality tests planning tool, virtual quality test result database, master data."}}, "f9d2322a-f237-46b4-939a-b90f81d0e922": {"node_ids": ["517ceb00-a21c-43fd-a3e1-999e2a837bc8", "32613ead-e644-4000-97ff-2bcad3244831"], "metadata": {"page_label": "484", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the current architecture of master data management in the DevManager system, and what are the key components involved in it?\n2. What are the target architecture components of master data management in the DevManager system, and how do they differ from the current architecture?\n3. What are the successor relationships between the various components of master data management in the DevManager system, and how do they relate to each other in the transformation model?", "prev_section_summary": "The section discusses the current gaps in an organization's IT landscape in terms of data management and planning tools. The proposed target architecture aims to address these gaps by unifying the functionality of different applications and using a development master data management (DMDM) system to manage master data and quality tests. The specific tools and systems that will be used in the target architecture include the DMDM system, a planning tool that includes planning for the product and quality tests, and a quality test assistance and result management tool. The section also mentions the existence of different applications for modifying products and managing quality tests.", "section_summary": "The section discusses the current and target architectures of master data management in the DevManager system, as well as the successor relationships between the various components of master data management in the transformation model. The section also mentions the key components involved in the current architecture, which include DevManager, Product Planning Tool, Quality tests planning tool, Physical quality test assistance tool, Physical quality test result database, Virtual quality test result database, Product class A assistance database, Product class B assistance database, QueryDev v1, MasterData v1 and v2. The target architecture includes Development master data management system (DMDM), Product and Quality test planning tool, Quality test assistance and result management tool, Product modification assistance database, MasterData v3 and PlanningData v1. The section also mentions the successor relationships between the various components of master data management in the transformation model, including the successor relationships between the services.", "excerpt_keywords": "1. Master data management, 2. Architecture, 3. Gap analysis, 4. Development master data management system (DMDM), 5. Product planning tool, 6. Quality tests planning tool, 7. Physical quality test assistance tool, 8. Virtual quality test result database, 9. Product modification assistance database, 10. Successor relationships."}}, "0b6c8e68-14a4-482b-a5a8-0e70a36ba52d": {"node_ids": ["41f6e3be-c7e8-483d-a345-4b16404411de", "f6bf074d-9de7-49e9-81be-78ff8f27d9b1"], "metadata": {"page_label": "485", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the document \"From Gaps to Transformation Paths in Enterprise Architecture Planning\"?\n2. What is the process for deriving gaps between the models of a current and target architecture for planning purposes?\n3. How does the action repository aid in the creation of possible transformation paths?", "prev_section_summary": "The section discusses the successor relationships between various tools and databases in the context of enterprise architecture planning. The successor relationships are used to determine the order in which these tools and databases are used in the planning process. The section also mentions the transformation model, which contains information about the successor and predecessor relationships between various services. The action repository is used to develop these services in the correct order. The key topics and entities discussed in the section include Product planning tool, Quality tests planning tool, Physical quality test assistance tool, Physical quality test result database, Virtual quality test result database, Product modification assistance database, Product class A assistance database, Product class B assistance database, MasterData v1, MasterData v2, MasterData v3, QueryDev v1, PlanningData v1, and the transformation model.", "section_summary": "The section discusses a document titled \"From Gaps to Transformation Paths in Enterprise Architecture Planning\" which outlines a process for deriving gaps between the models of a current and target architecture for planning purposes. The process involves using a set theoretic description to reuse existing information and aid in the detailing of the model of the target architecture. An action repository is also used to create possible transformation paths, which are sequences of actions. The solution considers a domain expert as an important part of the activities and assists her in the decision making process. The document also discusses limitations and requirements regarding the meta-model.", "excerpt_keywords": "Enterprise Architecture Planning, MasterData v3, dependencies, successor relationships, gap analysis, transformation model, action repository, business building blocks, meta-model, EAM approach, applications, services."}}, "87f7dab9-8506-4d0c-9b70-4cbc0003cd75": {"node_ids": ["e12b8d47-e893-40c2-af6f-cf52cf874035", "63c50146-9ae5-4fbf-bd1f-92f1f0bc0255"], "metadata": {"page_label": "486", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using a graph transformation approach for creating transformation paths in enterprise architecture planning?\n2. What is the difference between using SPARQL and a more sophisticated graph transformation approach for detailing the model of the target architecture?\n3. What is the proposed technical realization for modeling (M2M) transformation of ontologies to a model interpretable for a graph transformation approach?", "prev_section_summary": "The section discusses the limitations of the current EAM approach in addressing the concerns of stakeholders related to applications and their dependencies on services. It proposes the use of semantic web technologies for formalizing information sources in the context of EA planning, which has advantages such as a formal, unambiguous model and the possibility of reasoning and consistency checking. The section also mentions the importance of applications and their provided services in EA and the need for transformation paths that take technology architecture aspects into account.", "section_summary": "The section discusses the use of a graph transformation approach for creating transformation paths in enterprise architecture planning. It compares the usage of SPARQL and a more sophisticated graph transformation approach for detailing the model of the target architecture. The proposed technical realization for modeling (M2M) transformation of ontologies to a model interpretable for a graph transformation approach is discussed, along with related work in the field of enterprise architecture management. The section mentions the use of GROOVE as a mature graph transformation tool and the proposed solution requires a model to model (M2M) transformation of the ontologies to a model which is interpretable for a graph transformation approach.", "excerpt_keywords": "Enterprise Architecture Management, gap analysis, ontologies, graph transformation, Rule Interchange Format, GROOVE, modeling and modeling transformation, SPARQL, EA models, target architecture, transformation paths, ontology transformation, interpretable graph transformation, technical realization, related work, University of Oldenburg, Quasar Enterprise approach, technical standard, The Open Group."}}, "672b007c-4b9d-4463-a5d6-c4e1be707ef7": {"node_ids": ["117b0747-5d77-4e73-a31a-003d96ed1e86", "89b8dfba-f448-47bf-9f08-711337807584"], "metadata": {"page_label": "487", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What are the limitations of using gap analysis for selecting actions in Enterprise Architecture Planning?\n2. How does the Strategic IT Management approach use gap analysis to detect differences between the current and target architecture?\n3. What are the possible actions that can be considered to close gaps identified through gap analysis in Enterprise Architecture Planning?", "prev_section_summary": "The section discusses the gap analysis approach presented by the Institute for Information Technology at the University of Oldenburg in relation to the Quasar Enterprise app-roach. The approach is tightly coupled to the Quasar Enterprise app-roach and involves modeling the current and ideal application landscapes to generate a list of actions that would result in the ideal landscape. The section also identifies other approaches to gap analysis, including an approach from the University of Oldenburg and a technical standard from The Open Group. The Open Group technical standard is mentioned as being related to the gap analysis approaches discussed in the technical report.", "section_summary": "The section discusses the limitations of using gap analysis for selecting actions in Enterprise Architecture Planning (EAP). Gap analysis requires a detailed level of description when modeling both landscapes, and the data necessary to perform gap analysis on the entire application landscape on a detailed level considering operations is overwhelming. The Strategic IT Management approach uses gap analysis to detect differences between the current and target architecture, and possible actions to close the gaps identified through gap analysis are considered. The actions range from introducing a new application, adding or reducing functionality of an existing application, changing or adding services to the shut down of applications and services. The section also mentions the ArchiMate Implementation and Migration Extension, which introduces new concepts and modeling capabilities for EAP.", "excerpt_keywords": "1. ArchiMate\n2. Implementation\n3. Migration\n4. Extension\n5. Architecture\n6. Planning\n7. Gap analysis\n8. Process support maps\n9. Services\n10. Information objects"}}, "ca0fa447-93f1-41a7-b071-603606dd300e": {"node_ids": ["a9fada8e-463e-4063-8ce5-6f47fc12f666", "74c41a08-73c9-4d8e-8c3f-aa24daa056f7"], "metadata": {"page_label": "488", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the paper \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\" and how does it relate to transition architecture creation and domain expert support?\n2. How does the value-based weighting methodology in the paper support a domain expert in determining which transformation paths are more promising?\n3. What are the limitations and future work areas identified in the paper, and how do they relate to related work in the field of enterprise architecture planning?", "prev_section_summary": "The section discusses the purpose of conducting a gap analysis in enterprise architecture planning, the limitations of using ArchiMate for creating detailed target architectures, and how actions can be used to support the creation of transition architectures as plateaus between the current and target architecture. The section also introduces the concept of a gap in ArchiMate and its role in linking elements of two EA models. The section concludes with a discussion of future work in identifying situations where actions are necessary to create transition architectures as plateaus.", "section_summary": "The section discusses a paper titled \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation\" and its relevance to transition architecture creation and domain expert support. The paper presents a value-based weighting methodology to rank transformation paths based on different factors relevant for transformation planning. The action repository is also extended to cope with different EA models and concerns. The authors show how it is possible to get from identified gaps to transformation paths by creating a transformation model, detailing a target architecture, and using an action repository to create possible sequences of actions for transformation paths. They present a use case for parts of an application architecture and discuss the results and limitations of the solution, as well as future work to be addressed. The section references related work in the field of enterprise architecture planning, including ISO/IEC 42010, Winter and Fischer's \"Essential layers, artifacts, and dependencies of enterprise architecture,\" and Pulkkinen and Hirvonen's \"EA planning, development, and management process for agile enterprise development.\"", "excerpt_keywords": "1. Transformation architecture\n2. Action repository\n3. Target architecture\n4. Enterprise architecture\n5. Software engineering\n6. Systems design\n7. Architectural description\n8. Technical realization\n9. Agile enterprise development\n10. Systemic management"}}, "2463fbe5-c01e-4300-b6ec-eca629caa29b": {"node_ids": ["cbf994fe-1a71-4fea-bf84-67df6b4bab6e", "14415354-caa9-443b-9f88-a0e545afff92", "4c21dca7-54f7-4706-a32a-350f74581f7e"], "metadata": {"page_label": "489", "file_name": "From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "file_path": "docs\\From Gaps to Transformation Paths in Enterprise Architecture Planning.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the state-of-the-art in enterprise architecture management literature, and how has it evolved over time?\n2. How can enterprise models be applied to facilitate enterprise transformation, and what are some examples of their application?\n3. What is the role of information models in capturing the managed evolution of application landscapes, and how have they been used in this context?", "prev_section_summary": "The section discusses various topics related to enterprise architecture planning, including agile enterprise development, management of architectural decisions, complexity levels of representing dynamics, and IT governance. The authors and their works mentioned in the section include Pulkkinen, Hirvonen, Niemann, Aier, Gleichauf, Saat, and Winter. The section also mentions conferences and workshops where these authors presented their work, such as EDOCW'06 and HICSS'06.", "section_summary": "The section discusses the state-of-the-art in enterprise architecture management literature and its evolution over time. It also explores the application of enterprise models in facilitating enterprise transformation, and the role of information models in capturing the managed evolution of application landscapes. The section also mentions specific examples and tools used in these contexts, such as OWL web ontology language, SPARQL query language, and RDF Primer. Additionally, the section touches on semantic business process modeling and embedded EL+reasoning on programmable logic controllers.", "excerpt_keywords": "Enterprise architecture, management, transformation, models, information models, application landscapes, evolution, strategic management, IT, semantic web, web ontology, SPARQL, RDF, business process modeling, embedded reasoning, programmable logic controllers."}}, "619d0736-2192-4da9-828a-f33ef9a047db": {"node_ids": ["6f0b9627-227c-4eb5-92d1-4a57e4af6736", "bea362d2-325d-497a-97d2-019c0ffa3e61", "13b2c56c-d4b4-4173-b755-8b9800d99c3d", "84388c23-78e7-4d73-9103-a443e9587e62"], "metadata": {"page_label": "1", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the Heterogeneous Graph Transformer (HGT) architecture and how does it differ from other graph neural network (GNN) architectures?\n2. How does the HGT architecture handle dynamic heterogeneous graphs and Web-scale graph data?\n3. What is the performance of the HGT model compared to other state-of-the-art GNN baselines on various downstream tasks?", "prev_section_summary": "The section discusses the relationship between graph transformation and AI planning in enterprise architecture management. It also explores how enterprise topologies can be improved through segmentation, graph transformation, and analysis strategies. The role of graph grammars and computing by graph transformation in enterprise architecture management is also highlighted. The section includes references to various publications and research papers on the topic.", "section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) architecture, which is designed for modeling heterogeneous graphs. The architecture is different from other graph neural network (GNN) architectures in that it handles heterogeneity by designing node- and edge-type dependent parameters. The HGT architecture also includes the relative temporal encoding technique to capture dynamic structural dependency and the heterogeneous mini-batch graph sampling algorithm (HGSampling) for efficient and scalable training. The section presents experimental results showing that the proposed HGT model outperforms state-of-the-art GNN baselines on various downstream tasks. The dataset and source code of HGT are publicly available. Key entities discussed in the section include graph neural networks, heterogeneous information networks, representation learning, graph embedding, and graph attention.", "excerpt_keywords": "Graph Neural Networks, Heterogeneous Information Networks, Representation Learning, Graph Embedding, Graph Attention, Web-scale Graph Data, Dynamic Heterogeneous Graphs, Relative Temporal Encoding, Heterogeneous Mini-batch Graph Sampling, Open Academic Graph."}}, "a3e1e120-6732-4b4a-89ee-477130998e0e": {"node_ids": ["36c8ac35-d0df-4e32-888a-72bc36a252d2", "e244e139-e9a1-4e34-92b2-5835ada25da6", "bc9e8ddd-384f-4f71-b00b-acff631de8d2", "ade4f093-5318-40d7-9b81-2204d26fff8e", "8e1b103b-1d79-485b-8bbe-3161ba6db37f"], "metadata": {"page_label": "2", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the main goal of the proposed Heterogeneous Graph Transformer (HGT) architecture?\n2. How does the node- and edge-type dependent attention mechanism in HGT handle graph heterogeneity?\n3. What is the relative temporal encoding (RTE) strategy proposed to handle graph dynamics in HGT?", "prev_section_summary": "The section discusses the limitations of existing heterogeneous graph neural network models and introduces the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation. The model addresses the limitations by accounting for the dynamic nature of heterogeneous graphs, capturing unique properties of each type of heterogeneous graph, and modeling Web-scale heterogeneous graphs. The section uses the OAG as an example to illustrate the limitations of existing models and how the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation can capture its unique properties.", "section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) architecture, which is proposed to handle the limitations and challenges of existing heterogeneous graph neural networks (GNNs) in dealing with large-scale heterogeneous graphs. The HGT architecture introduces a node- and edge-type dependent attention mechanism to maintain specific representation spaces for nodes and edges of different types, while allowing connected nodes in different types to interact and aggregate messages. The architecture also incorporates information from high-order neighbors of different types through message passing across layers, and proposes a relative temporal encoding (RTE) strategy to handle graph dynamics. The section highlights the benefits of the HGT architecture, including its ability to capture network dynamics, avoid customized meta paths, and be scalable to Web-scale graphs.", "excerpt_keywords": "1. Heterogeneous Graph Neural Networks\n2. Node- and Edge-type Dependent Representations\n3. Graph Dynamics\n4. Scalability\n5. Heterogeneous Mutual Attention\n6. Meta Relations\n7. Attention Mechanism\n8. Message Passing\n9. Relative Temporal Encoding\n10. Web-scale Graphs"}}, "1dd47884-7047-44eb-8ed3-82f98752a400": {"node_ids": ["3760465e-0f22-4a23-9bef-025caa471241", "4bf4e7f0-65c0-4f05-a1ee-fd3e3a61809b", "1354b19a-c9b9-44fe-825b-246ecf80ea48", "00a98235-a28f-4d0d-8435-9396a1e73830", "b5b9b8c1-06ce-4c13-8251-cbc1faf88485"], "metadata": {"page_label": "3", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the general GNN framework and what are the most important GNN operators?\n2. What are some examples of homogeneous GNN architectures and how do they differ from each other?\n3. What is the difference between GCN, GraphSAGE, and GAT, and how do they differ in their aggregation operations?", "prev_section_summary": "The section discusses heterogeneous graphs, which are graphs that contain nodes and edges of different types. The section explains the difference between static and dynamic heterogeneous graphs and how the dynamic perspective of heterogeneous graphs can be explored and studied. The section also discusses recent lines of research on mining heterogeneous graphs, including node classification, clustering, ranking, and representation learning. The section concludes by discussing the success of graph neural networks for relational data.", "section_summary": "The section discusses the general Graph Neural Network (GNN) framework and its most important operators, Extract and Aggregate. It also describes some examples of homogeneous GNN architectures, including Graph Convolutional Network (GCN), GraphSAGE, and Graph Attention Network (GAT), and their differences in aggregation operations. The section also mentions the Heterogeneous Graph Transformer, a WWW '20 conference paper that discusses message passing in GNNs.", "excerpt_keywords": "GNN, message passing, node representation, GNN framework, Extract, Aggregate, GCN, GraphSAGE, GAT, attention mechanism, neighborhood information, graph convolutional network, linear projection, non-linear activation, RNN unit, sum, max, pooling, normalization, utility, fairness, positivity, heterogeneous graph transformer, WWW '20, Taipei, Taiwan."}}, "56f63095-98b7-46bf-a7dc-790d001587c3": {"node_ids": ["96e72751-1642-4479-9d54-b5a803ecee88", "2d8ae984-b0ca-4c47-b83c-b9c253d2a67f", "c1eb1480-19ca-4c60-aced-2a66c976cc36", "93fc5a11-ba77-4314-a9aa-69af38413333"], "metadata": {"page_label": "4", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the overall architecture of the Heterogeneous Graph Transformer model and how does it work?\n2. How does the Heterogeneous Graph Transformer model handle heterogeneous graphs with different node types and their corresponding meta relations?\n3. What are the three basic operators used in the Heterogeneous Graph Transformer model and how do they work?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) model, which is a multi-modal graph neural network used for corporate relative valuation. The model consists of three components: Heterogeneous Mutual Attention, Heterogeneous Message Passing, and Target-Specific Aggregation. The HGT model calculates the mutual attention between source and target nodes, and uses this information to aggregate messages from the source nodes to the target node. The model can be stacked to obtain node representations of the whole graph, which can be used for end-to-end training or fed into downstream tasks. The section also provides a brief introduction to attention-based graph neural networks.", "section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) model, which is a multi-modal graph neural network used for corporate relative valuation. The model works by taking a sampled heterogeneous sub-graph with a target node, source nodes, and their corresponding meta relations as input and learning contextualized representations for each node. The HGT model includes three components: meta relation-aware heterogeneous mutual attention, heterogeneous message passing from source nodes, and target-specific heterogeneous message aggregation. The three basic operators used in the HGT model are Attention, Message, and Aggregate. The Graph Attention Network (GAT) is mentioned as an example of a model that uses the Attention operator, but it is noted that GAT assumes that all nodes have the same feature distributions, which is usually incorrect for heterogeneous graphs.", "excerpt_keywords": "1. Heterogeneous Graph Transformer, 2. Graph Attention Network, 3. Meta relation-aware heterogeneous mutual attention, 4. Heterogeneous message passing, 5. Target-specific heterogeneous message aggregation, 6. Attention, 7. Message, 8. Aggregate, 9. Node type, 10. HGT"}}, "142f0092-e09f-411b-81fc-dd5b2fe76ffb": {"node_ids": ["c46ef647-4874-4fbc-be8f-80e3c1b4dcd5", "eccc65c3-7dd9-4b2d-a4ef-4e495f0c90c9", "fc8cafcd-884b-4f2c-868b-21b9b442fd04", "f271ba5c-b90d-41b8-ba61-abd944580ccb"], "metadata": {"page_label": "5", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Graph Transformer model in the context of corporate relative valuation?\n2. How does the model incorporate meta-relations of edges into the attention and message passing processes?\n3. What is the role of the prior tensor \u00b5 in the Heterogeneous Graph Transformer model?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer model, which is a type of multi-modal graph neural network used for corporate relative valuation. The model uses attention heads and vector dimensions to capture different semantic relations between node types in heterogeneous graphs. The WATT \u03d5(e) matrix is also discussed, which is used to capture different semantic relations between node types. The model is designed to capture the unique characteristics of heterogeneous graphs, such as different edge types between node pairs.", "section_summary": "The section discusses the Heterogeneous Graph Transformer model, which is used for corporate relative valuation. The model incorporates meta-relations of edges into the attention and message passing processes. The prior tensor \u00b5 is used to denote the general significance of each meta-relation triplet and serves as an adaptive scaling to the attention. The section also explains the Heterogeneous Message Passing process, which incorporates the meta-relations of edges into the message passing process to alleviate the distribution differences of nodes and edges of different types.", "excerpt_keywords": "1. Heterogeneous Graph Transformer, 2. Attention, 3. Message Passing, 4. Meta Relations, 5. Node Pairs, 6. Target Nodes, 7. Adaptive Scaling, 8. Softmax, 9. Multi-Head Message, 10. Edge Dependency"}}, "a6390232-e74a-4453-ac1c-f469dac6f577": {"node_ids": ["6abd790d-8794-4c04-bcf0-f669a2529da4", "b42e822c-3234-4543-99f4-28f3def94b28", "6a2ab275-5181-4cc2-b554-cc30e92ebb1d", "f4110556-56cf-4a3b-854c-63133116eb05", "09cb40e9-ec9e-400f-9eef-1708cefdcd04"], "metadata": {"page_label": "6", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of using sinusoid functions as a basis for the temporal encoding in Heterogeneous Graph Transformer (HGT)?\n2. How does the Heterogeneous Mini-Batch Graph Sampling algorithm (HGSampling) address the scalability issue of full-batch GNN training for Web-scale heterogeneous graphs?\n3. What is the inductive timestamp assignment method used in HGT for training Web-scale heterogeneous graphs with dynamic information?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) model for modeling heterogeneous graphs. HGT differs from existing models by using triplet parameterization for parameter sharing, which helps in achieving faster adaptation and generalization for relations with few occurrences while still maintaining the specific characteristics of different relationships' operators. The section also introduces the Relative Temporal Encoding (RTE) technique for handling dynamic information in heterogeneous graphs, which is inspired by Transformer's positional encoding method. RTE allows nodes and edges with different timestamps to interact with each other and model the dynamic dependencies in heterogeneous graphs.", "section_summary": "The section discusses the use of sinusoid functions as a basis for temporal encoding in Heterogeneous Graph Transformer (HGT) and the Heterogeneous Mini-Batch Graph Sampling algorithm (HGSampling) for scalable training of Web-scale heterogeneous graphs. The inductive timestamp assignment method is also introduced for training Web-scale heterogeneous graphs with dynamic information. The section highlights the challenges of full-batch GNN training for Web-scale graphs and the need for scalable and efficient methods for training heterogeneous graphs.", "excerpt_keywords": "1. Graph Neural Networks (GNNs), 2. Heterogeneous Graphs, 3. Web-scale, 4. Dynamic Information, 5. Temporal Encoding, 6. Target Node, 7. Source Node, 8. Timestamp Assignment, 9. Mini-Batch Graph Sampling, 10. Inductive Learning"}}, "bb57f321-c60f-4e25-8444-a5cdbee87b27": {"node_ids": ["004441c6-45e6-4480-93d6-4f85f41c09d3", "7ea36911-fd07-4f09-b748-1cb9ca04f495", "20458bfb-7830-4b2e-a62f-8e15a800029c"], "metadata": {"page_label": "7", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Graph Transformer algorithm and how does it work?\n2. How does the Add-In-Budget algorithm work in the context of Heterogeneous Graph Transformer?\n3. What is the role of timestamp assignment in the Heterogeneous Graph Transformer algorithm and how does it work?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer model, which is used for training Graph Neural Networks (GNNs) on Web-scale heterogeneous graphs. The algorithm used to sample a sub-graph from the initial nodes in the model is described, as well as the purpose of assigning timestamps to nodes in real-world heterogeneous graphs. The section also explains the concept of plain nodes, which are nodes that are not associated with a fixed time, and the need for inductive timestamp assignment to decide which timestamp(s) to attach to these nodes.", "section_summary": "The section discusses the Heterogeneous Graph Transformer algorithm, which is used for corporate relative valuation. The algorithm works by sampling nodes from a heterogeneous graph and assigning them timestamps based on event nodes they are linked with. The Add-In-Budget algorithm is also introduced, which updates the budget for each node type based on the normalized degree of the added node t. The section also mentions the role of timestamp assignment in the Heterogeneous Graph Transformer algorithm and how it works.", "excerpt_keywords": "1. Heterogeneous Graphs\n2. Timestamp Assignment\n3. Inductive Timestamping\n4. Event Nodes\n5. Publication Behavior\n6. Publication Date\n7. Plan Nodes\n8. Temporal Dependency\n9. Normalized Degree\n10. Budget Allocation"}}, "70efd5d5-bcda-4698-a7e4-c0ba91bf514a": {"node_ids": ["13bf3d72-249e-4505-8136-7246fad2c783", "f7527283-444c-483e-857d-e02388c28f30", "4b37021e-9968-4f85-b8b9-3cb8e730ad15", "36bb2f32-dc7b-4109-94a8-ec50e301c6a7", "008368f5-ed54-4b4a-8e7e-f808ace630b7"], "metadata": {"page_label": "8", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation?\n2. How does the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation differentiate between the different node types and edge types in the Open Academic Graph (OAG)?\n3. What are the performance metrics used to evaluate the effectiveness of the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation on the Open Academic Graph (OAG)?", "prev_section_summary": "The section discusses the purpose of constructing domain-specific subgraphs from OAG for testing the generalization of a proposed heterogeneous graph neural network model. The size and distinguishability of the Computer Science (CS) and Medicine (Med) academic graphs are compared to other commonly used academic datasets in existing heterogeneous GNN and graph mining studies. The dataset and code for the proposed heterogeneous graph neural network model are publicly available at https://github.com/acbull/pyHGT.", "section_summary": "The section discusses the Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation, which is a graph neural network that is used to predict the relative valuation of a company based on various factors such as its papers, authors, fields, venues, and institutions. The network differentiates between the different node types and edge types in the Open Academic Graph (OAG), which is a dataset that contains information about academic papers, authors, fields, venues, and institutions. The performance metrics used to evaluate the effectiveness of the network on the OAG include precision, recall, F1 score, and accuracy.", "excerpt_keywords": "1. Open Academic Graph (OAG)\n2. Node types: Paper, Author, Field, Venue, Institute\n3. Field levels: L0 to L5\n4. Author orders: first, last, others\n5. Venue types: journal, conference, preprint\n6. Self-loop connection\n7. GNN architectures\n8. Keyword extraction\n9. Information retrieval\n10. Academic network analysis"}}, "7ee672fa-2873-49ed-bc6c-25464d3d468c": {"node_ids": ["e2dd3028-2df8-4f0c-9631-86451db2c111", "67058179-72da-4f35-adfc-ccbaa8300641", "a21bd965-501e-42d3-b74b-c49995ad5f4f", "8054dd11-78a3-4777-b1ea-55132e226d1a", "89e86816-c6c5-42d5-bc8b-ab254044ee5a", "ceab7ae5-1fbe-4481-8356-475bf9774c29"], "metadata": {"page_label": "9", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the performance of Heterogeneous Graph Transformer in terms of NDCG and MRR for the CSPaper\u2013Field (L1) task compared to other GNN models such as GCN, RGCN, GAT, HetGNN, HAN, HGT-RT, and HeterHGT-RT?\n2. How does the batch time of Heterogeneous Graph Transformer compare to other GNN models for the CSPaper\u2013Field (L1) task, and what is the impact of batch time on the performance of these models?\n3. What is the difference in performance between Heterogeneous Graph Transformer and other GNN models such as GCN, RGCN, GAT, HetGNN, HAN, HGT-RT, and HeterHGT-RT for the Paper\u2013VenueNDCG task, and what factors contribute to this difference?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT) model, which is used for representing the title of each paper and handling heterogeneous network structures in the field, venue, and institute nodes. The model uses a pre-trained XLNet model to get the representation of each word in the title and an adaptation layer to make a fair comparison with homogeneous GNN baselines. The implementation details of the model are also discussed, including the use of 256 as the hidden dimension and 8 as the number of heads in the multi-head attention-based methods.", "section_summary": "The section discusses the performance of Heterogeneous Graph Transformer (HGT) in terms of NDCG and MRR for the CSPaper\u2013Field (L1) task compared to other GNN models such as GCN, RGCN, GAT, HetGNN, HAN, HGT-RT, and HeterHGT-RT. It also compares the batch time of HGT to other GNN models for the CSPaper\u2013Field (L1) task and discusses the difference in performance between HGT and other GNN models for the Paper\u2013VenueNDCG task. The section also mentions the number of parameters and batch time of each model, as well as the factors that contribute to the difference in performance between HGT and other GNN models for the Paper\u2013VenueNDCG task.", "excerpt_keywords": "GNN Models, GCN, RGCN, GAT, HetGNN, HAN, HGT-RT, E-HeterHGT+RT, E-HeterHGT-RT, E-HeterHGT+RT, NDCG, MRR, Paper-Field, L1, Paper-Field, L2, Paper-Venue, NDCG, MRR, Paper-Venue."}}, "29d3becc-38ca-47cb-9332-674e4b7909a6": {"node_ids": ["cbfd04f2-7520-418b-bd69-a2200a9c66ed", "ec2fabc6-1aaf-4704-85b0-f81e974ffb56", "9ec2d462-902a-4147-98f1-ce3d1c978d93", "68b2b3e6-15b0-4bfc-b376-00288ff26c68", "f9877c88-0e08-4067-b0ee-40dac187596f"], "metadata": {"page_label": "10", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the significance of using Relative Temporal Encoding in Heterogeneous Graph Transformer (HGT) for capturing graph dynamics?\n2. How does the removal of heterogeneous weight parameterization and Relative Temporal Encoding affect the performance of HGT?\n3. Can using meta relations to parameterize weight matrices achieve competitive performance with fewer resources compared to a baseline that keeps a unique weight matrix for each relation?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT), a graph neural network that improves the performance of graph neural networks by modeling heterogeneous edges according to their meta relation schema. HGT achieves relative performance gains of 15-19% in terms of NDCG and 18-21% in terms of MRR compared to other heterogeneous graph neural network baselines. HGT outperforms GCN, GAT, RGCN, HetGNN, and HAN by 20% for the four tasks on all three large-scale datasets, while having fewer parameters and comparable batch time.", "section_summary": "The section discusses the significance of using Relative Temporal Encoding in Heterogeneous Graph Transformer (HGT) for capturing graph dynamics. An ablation study is conducted to analyze the effects of removing heterogeneous weight parameterization and Relative Temporal Encoding on the performance of HGT. The results show that parameterizing with meta relations and using Relative Temporal Encoding are significant for achieving competitive performance with fewer resources. A case study is also conducted to evaluate how Relative Temporal Encoding can help HGT to capture graph dynamics by showing the evolution of conference topics.", "excerpt_keywords": "Heterogeneous weight parameterization, Relative Temporal Encoding, ablation study, graph dynamics, conference topic, evolution, WWW, KDD, NeurIPS, timestamps, sub-graphs, representations, euclidean distances, similarity, relationships, change, keywords."}}, "f3ccaf27-46f6-4aaa-b10d-195f2156f4b6": {"node_ids": ["8a448c50-08b1-47f8-853b-919396983891", "96dc6fbf-2f14-4e07-ac4c-0f3385813d2b", "15062f1b-c7f6-45d4-8ae4-7fedc8eda488", "4c5cb8e7-665e-410f-920d-33413ad7c183", "f9034dee-5640-4342-9a2d-c72ad0780d4c"], "metadata": {"page_label": "11", "file_name": "Heterogeneous Graph Transformer.pdf", "file_path": "docs\\Heterogeneous Graph Transformer.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the purpose of the Heterogeneous Graph Transformer paper presented at the WWW \u201920 conference?\n2. How does the FastGCN algorithm improve the efficiency of learning with Graph Convolutional Networks?\n3. What is the significance of the Stochastic Training of Graph Convolu-tional Networks with Variance Reduction technique in the field of graph neural networks?", "prev_section_summary": "The section discusses the Heterogeneous Graph Transformer (HGT), a type of Graph Neural Network (GNN) that is designed to handle heterogeneous graphs with multiple types of nodes and edges. HGT outperforms other GNN base-lines on various downstream tasks and has the potential to generate heterogeneous graphs, such as predicting new papers and their titles. The development of HGT was supported by funding sources including the National Science Foundation (NSF), Okawa Foundation, and Amazon Research Award. The key contributors to the development of HGT include Xiaodong Liu.", "section_summary": "The section discusses the Heterogeneous Graph Transformer paper presented at the WWW '20 conference, which focuses on developing a multi-modal graph neural network for corporate relative valuation. The paper also mentions the FastGCN algorithm, which improves the efficiency of learning with Graph Convolutional Networks, and the Stochastic Training of Graph Convolu-tional Networks with Variance Reduction technique, which is significant in the field of graph neural networks. Additionally, the section references several other papers related to graph neural networks and representation learning.", "excerpt_keywords": "Graph convolutional networks, importance sampling, stochastic training, heterogeneous networks, metapath2vec, globalization, scientific collaborations, citations, innovations, neural message passing, quantum chemistry, inductive representation learning, deep residual learning, image recognition, semi-supervised classification."}}, "fb80ecac-79cd-47f7-a382-727a1c8a533e": {"node_ids": ["6c789761-7721-438b-83f6-32580146d47a", "5243f89c-b9ab-442b-b607-6b8ea6277c0f", "ee5db39c-b21e-4726-9736-122a98b25b66", "2b8a9431-524f-4944-8ab6-cc5042aed4e8"], "metadata": {"page_label": "93", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the relationship between network-oriented modeling and the dynamical systems perspective on mental processes?\n2. Can any adaptive dynamical system be modeled by a self-modeling network?\n3. What are the applications of a network-oriented modeling approach in scientific domains?", "prev_section_summary": "The section discusses the difference between Heterogeneous Graph Neural Network (HGNN) and OAG in linking large-scale heterogeneous entity graphs. It also explains how Layer-Dependent Importance Sampling improves the training of deep and large graph convolutional networks. The significance of Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation in the field of graph neural networks is also highlighted. The authors of the section are Chuxu Zhang, Dongjin Song, Chao Huang, Ananthram Swami, and Nitesh V. Chawla, Fanjin Zhang, Xiao Liu, Jie Tang, Yuxiao Dong, Peiran Yao, Jie Zhang, Xiaotao Gu, Yan Wang, Bin Shao, Rui Li, and Kuansan Wang. The document was created on 2023-11-06, last modified on 2023-10-20, and last accessed on 2023-11-30. The title of the document is \"Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative", "section_summary": "The section discusses the relationship between network-oriented modeling and the dynamical systems perspective on mental processes. It is mathematically proven that any dynamical system can be modeled as a temporal-causal network model and that any adaptive dynamical system (of any order) can be modeled by a self-modeling network (of the same order). The scope of applicability of a network-oriented modeling approach covers mental processes described by networks, social (interaction) processes described by social network models, and more. The section also mentions the longstanding tradition of network-oriented modeling in AI.", "excerpt_keywords": "Dynamical system, Adaptive, Network model, Self-modeling network, Mathematical analysis, Temporal-causal network, Causal pathways, Causal explanation, AI, Scientific domains."}}, "74571bde-6fcb-4af9-ab8a-35c53b325152": {"node_ids": ["a42737a5-ac14-4639-b0a9-050b2bc5b0c2", "b8e3088d-2326-4b7c-b6e7-6027432a2844", "b7294d1e-6da7-4f44-ad24-01f6c545c93f", "d5e7c57b-e43f-4036-a5db-b580d97735d2", "fa64ff21-9041-4fa6-bf55-b326d25811e8", "29fc01a6-e599-43ea-bcc1-2787e7698c86", "34f3377d-b7e2-4dd9-917b-b5cd9adcfcb7"], "metadata": {"page_label": "94", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the definition of a state-determined system and how does it relate to the notion of a dynamical system?\n2. What are the three features of a state-determined system and how are they reflected in modeling and simulation?\n3. How does the state-determined system assumption relate to the notions of temporal factorisation and criterial causation?", "prev_section_summary": "The section discusses the relationship between temporal-causal network models and adaptive dynamical systems in modeling any set of first-order equations. The authors show mathematically that temporal-causal network models can model any dynamical system or any set of first-order differential equations. They also show how any adaptive dynamical system can be modeled by a self-modeling network model. The authors discuss the notion of state-determined system from Ashby (1960) and Port and Van Gelder (1995) and show that it is equivalent to a set of first-order differential equations. The authors also provide a mathematical analysis that shows how any adaptive dynamical system can be modeled by a self-modeling network model.", "section_summary": "The section discusses the concept of a state-determined system and its relation to the notion of a dynamical system. The state-determined system assumption is adopted from Ashby (1960) and taken as the basis to describe what a dynamical system is. Three features of a state-determined system are discussed: the future behavior cannot depend on states the system might have been in before the current state, the current state determines future behavior implies the existence of some rule of evolution describing the behavior of the system as a function of its current state, and future behaviors are uniquely determined. The section also relates the state-determined system assumption to the notions of temporal factorization and criterial causation.", "excerpt_keywords": "state-determined system, dynamical system, modeling, simulation, rule of evolution, differential equations, network models, temporal factorization, criterial causation, past history, current state, future behavior, uniqueness, state space sequences, fork, prediction, simulation, rule of evolution, specification, succinct, useful, Ashby, Port, van Gelder, Treur."}}, "1deea7fa-6cef-41a1-96fe-bcf0ad499d43": {"node_ids": ["572359a5-bce2-4e26-9bdb-9369ef6e0c84", "93b1b873-30ec-4f04-b951-d4dafe32d0c2", "96066384-e0d4-45fc-a5d7-3ff5958ec1fb", "4482903a-bf89-422e-be1f-cffd4bf551fe", "fd9baaf3-ca64-4a04-a18d-2e4d634048a2"], "metadata": {"page_label": "95", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the mathematical formula used to combine the output of multiple combination functions in a network model?\n2. How are the network characteristics of a model specified in the role matrices format?\n3. What is the difference between adaptive and non-adaptive network models, and how do they interact with each other?", "prev_section_summary": "The section discusses the use of heterogeneous multi-modal graph neural networks in designing network models and their dynamics in a declarative manner. The basic combination functions bcfi(..), i =1, .., m are used to support the design of network models for any application. Mathematical functions and relations are used to analyze adaptive dynamical systems and self-modeling network models in mental processes. The section also mentions the importance of understanding the states from which incoming connections are obtained.", "section_summary": "The section discusses the mathematical formula used to combine the output of multiple combination functions in a network model, the role matrices format used to specify the network characteristics of a model, and the difference between adaptive and non-adaptive network models. The section also explains how self-models can be used within a network to provide a network-oriented conceptualization of the dynamics of the network.", "excerpt_keywords": "1. Network model, 2. Network characteristics, 3. Combination function, 4. Role matrices format, 5. Connectivity characteristics, 6. Aggregation characteristics, 7. Timing characteristics, 8. Adaptive networks, 9. Co-evolution, 10. Self-models"}}, "5acd044b-9d8d-47f7-b88e-cf4ea4514f96": {"node_ids": ["c5fc4b53-c961-43ee-93a3-a051764fb0c3", "c7751fc1-42be-4936-81e6-4344d4449a51", "0a646399-e0de-4836-b25d-04ae232449c3", "07a0d979-1c2f-46c4-a0a7-91de295cfd65", "e85dffe1-a000-42dd-99a7-4c5eabf7ef14", "e9e11adf-a9e5-42a1-9eed-7b53e2289646"], "metadata": {"page_label": "96", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. How can a state-determined system be transformed into a format that more directly relates to causal relationships between states?\n2. What is the conceptual representation of a dynamical system as a causal graph, and how can it be visualized?\n3. How can a numerical representation of a dynamical system be transformed into a causal graph model, and what is the mathematical formula for this transformation?", "prev_section_summary": "The section discusses the use of first-order differential equations to describe the state-determined system in a given context. It also introduces different types of network characteristics and self-model states, and their roles in the system. The combination functions, weights, and parameters are also discussed and their significance in the context of self-modeling network models is highlighted. The section concludes with a summary of the key topics and entities.", "section_summary": "The section discusses the transformation of state-determined systems into a format that more directly relates to causal relationships between states. This can be achieved by representing the dynamical system as a causal graph, which can be visualized in a network model. The section also provides a mathematical formula for transforming a numerical representation of a dynamical system into a causal graph model. The key topics and entities discussed in the section include dynamical systems, causal relationships, graph neural networks, and differential equations.", "excerpt_keywords": "1. Dynamical systems\n2. Network models\n3. Causal relationships\n4. Graphical representation\n5. Temporal-causal network format\n6. State-determined systems\n7. State changes\n8. Subset of states\n9. Causal effects\n10. Differential equations"}}, "1a293364-9a65-4363-a78e-033057f53944": {"node_ids": ["67141005-9c62-489d-9374-bff1fa9da7ae", "59b16652-6ade-45b3-9ad5-6bbaebd78823", "9e946097-f0e9-444f-8e38-8402b039005c", "894b0c9d-4aac-46ea-9b88-5450d1f33057", "1d88f0c5-ed50-41e1-99f4-b6bc66e49c64"], "metadata": {"page_label": "97", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the mathematical relationship between adaptive dynamical systems and self-modeling network models?\n2. How can the temporal-causal network model described by the differential equations (8) be transformed into an instance of a more general temporal-causal model?\n3. What is the significance of the numerical representation (3) in the context of temporal-causal network models?", "prev_section_summary": "The section discusses the mathematical representation of combination functions used in differential equations related to adaptive dynamical systems and self-modeling network models. The combination functions are built from sum and product functions and are used to model the dynamics and adaptivity of mental processes. The section also provides a graphical network representation of the example model based on the given differential equation representation. The significance of the combination functions in the context of these models is not explicitly stated.", "section_summary": "The section discusses the relationship between adaptive dynamical systems and self-modeling network models, specifically in the context of temporal-causal network models. The section presents a numerical representation of a temporal-causal network model and explains how the differential equations representation can be transformed into an instance of the numerical representation of a more general temporal-causal model. The section also discusses the significance of the numerical representation in the context of temporal-causal network models.", "excerpt_keywords": "1. Temporal-causal network model\n2. Dynamical system model\n3. Numerical representation\n4. Differential equations\n5. Transformation\n6. Exact manner\n7. General temporal-causal model\n8. \u03b7Xi =1\n9. \u03c9XiXj =1\n10. Cognitive Systems Research"}}, "b0d761ef-9bec-46bf-a112-261d0f389c3a": {"node_ids": ["cb0ecf56-27a0-45f6-b6ea-9169066c6cb8", "39a7d3e7-ab49-4d23-8cef-4bd4f8a5b3b2", "e41dabeb-3e80-45bf-b38f-785a14c25b34", "f331d817-57db-4bd2-aedb-d960b44cf61e", "ce244400-592a-41ac-9992-aa96623a8a49"], "metadata": {"page_label": "98", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the mathematical model used to describe the dynamics of variables Xi in the lower level dynamical system?\n2. How do the parameters Pi,j become time-dependent in the upper level component of the model?\n3. What is the transformation applied to the base dynamical system (14) to obtain the combination functions hi(\u2026) found for the base level states Xi?", "prev_section_summary": "The section discusses the relationship between adaptive dynamical systems and self-modeling network models in mental processes, and how mathematical analysis can be used to understand the dynamics and adaptivity of mental processes. The role of role matrices in specifying the example set of differential equations in the context of mental processes is also discussed. The section is authored by J. Treur.", "section_summary": "The section discusses the mathematical model used to describe the dynamics of variables Xi in the lower level dynamical system, how the parameters Pi,j become time-dependent in the upper level component of the model, and how the transformation applied to the base dynamical system (14) obtains the combination functions hi(\u2026) found for the base level states Xi. Additionally, the section describes how a similar transformation is applied to the adaptation level described by (15) to define function qi,j(..). The key topics and entities discussed in this section include adaptive dynamical systems, self-modeling network models, mathematical analysis, heterogeneous multi-modal graph neural networks, and corporate relative valuation.", "excerpt_keywords": "1. Dynamical systems\n2. Parameters\n3. Lower level component\n4. Upper level component\n5. Time-dependent\n6. Differential equations\n7. Transformation\n8. Combination functions\n9. Adaptation level\n10. Pi,j"}}, "d64f1393-8ea3-4495-b60d-300797f050d1": {"node_ids": ["69ee04f9-a2b5-4f68-a89f-5e0a468ed6bb", "4fd74f11-08d6-41bd-a62d-3fe64bbaee4a", "3ee06420-3e90-4e62-bdc6-a81fd84a1271", "0b71a2c3-b090-4c63-9246-2c508d56a785", "99ac42b4-e3fc-40fc-9cc9-c28519b61054"], "metadata": {"page_label": "99", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the mathematical relationship between network-oriented modeling and the dynamical systems perspective on mental processes?\n2. Can any adaptive dynamical system be accurately modeled using a self-modeling network?\n3. What is the philosophical significance of the relationship between dynamical systems and network-oriented modeling in understanding mental processes?", "prev_section_summary": "The section discusses the mathematical analysis of adaptive dynamical systems and self-modeling network models. It explains the new combination functions cX6(..) to cX9(..) and their relationship to the adaptation system and base dynamical system in an adaptive dynamical system. The section also provides a graphical representation of the self-modeling network model based on the given differential equation representation of an adaptive dynamical system. The key topics and entities discussed in the section include adaptive dynamical systems, self-modeling network models, mathematical analysis, combination functions, and graphical representation.", "section_summary": "The section discusses the relationship between network-oriented modeling and the dynamical systems perspective on mental processes. It explains how any dynamical system can be modeled as a temporal-causal network model and any adaptive dynamical system can be modeled by a self-modeling network. The section also provides mathematical transformations for these models and discusses their philosophical significance in understanding mental processes. The key topics and entities discussed in the section include network-oriented modeling, dynamical systems perspective, mental processes, adaptive dynamical systems, temporal-causal network models, self-modeling networks, and mathematical transformations.", "excerpt_keywords": "dynamical systems, mental processes, network-oriented modeling, temporal-causal network model, adaptive dynamical systems, self-modeling network, mathematical transformations, exact modeling, philosophical context, dynamical systems perspective."}}, "d5c086c9-cde5-47c6-b8d9-3bd77a6a1b8c": {"node_ids": ["c87cd0d2-0b34-4099-88fd-ef65e4c7708a"], "metadata": {"page_label": "100", "file_name": "On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "file_path": "docs\\On the dynamics and adaptivity of mental processes Relating adaptive dynamical systems and self-modeling network models by mathematical analysis.pdf", "creation_date": "2023-11-06", "last_modified_date": "2023-10-20", "last_accessed_date": "2023-11-30", "document_title": "Heterogeneous Multi-Modal Graph Neural Network for Corporate Relative Valuation", "questions_this_excerpt_can_answer": "1. What is the relationship between adaptive dynamical systems and self-modeling network models in the context of mental processes?\n2. How can mathematical analysis be used to model the dynamics and adaptivity of mental processes?\n3. What is the role of higher-order adaptivity in network-oriented modeling for adaptive biological, mental, and social network models?", "prev_section_summary": "The section discusses the relationship between adaptive dynamical systems and self-modeling network models in cognitive modeling. It also explores the limitations of network models in modeling networks and how they can be improved. The section highlights the use of network-oriented modeling based on temporal-causal networks to address the complexity of cognitive, affective, and social interactions. The section also includes references to relevant research papers and the author's work on the topic.", "section_summary": "The section discusses the relationship between adaptive dynamical systems and self-modeling network models in the context of mental processes. Mathematical analysis is used to model the dynamics and adaptivity of mental processes. The role of higher-order adaptivity in network-oriented modeling for adaptive biological, mental, and social network models is also discussed. The section references several papers and articles that explore the topic in more detail.", "excerpt_keywords": "1. Network science\n2. Adaptive networks\n3. Higher-order adaptivity\n4. Cognitive systems\n5. Temporal factorisation\n6. Criterial causation\n7. Informational content\n8. Free will\n9. Dynamical approach\n10. Cognitive science"}}}, "docstore/metadata": {"5ef454ee-9d20-4491-ba57-d0d4aa559569": {"doc_hash": "afeb43fac3703a86579f4bf153211af428a6a322710dec9d1f8f87aa6305abf3", "ref_doc_id": "f40bb129-03c8-42b9-80ee-694879b4e4ae"}, "316f1e3e-adc4-4999-8696-fc25f0e20b50": {"doc_hash": "3a549764bd354d9c48b3f179a3b70a8dae56274e0c61b849b03d46b547d12f87", "ref_doc_id": "f40bb129-03c8-42b9-80ee-694879b4e4ae"}, "5c3e77ea-e060-4453-861a-af0a9c89a785": {"doc_hash": "5ad45c87f8f1e75d51a73ce03b2f728b5c1a24247d47197612a3bafbd518d714", "ref_doc_id": "f40bb129-03c8-42b9-80ee-694879b4e4ae"}, "cd77e6c2-e980-4dce-9375-a31aa31ccde4": {"doc_hash": "4b819f9d06a019f7a1599c839290e7c82d0b08c8049021be1e3afbe35436dc63", "ref_doc_id": "f40bb129-03c8-42b9-80ee-694879b4e4ae"}, "a11e5164-0951-41fc-9479-c9cccb09e5d8": {"doc_hash": "4f39a624741dbc098f2a19a86b333808c97b4c818ca58025855bb253becc9eec", "ref_doc_id": "f40bb129-03c8-42b9-80ee-694879b4e4ae"}, "175214df-164e-4f15-baa4-95cce9ade065": {"doc_hash": "3bfb07c09e85dbd9a8598da2b1a4c303d2f4614ecf9e4da33ebcc4487bc7307e", "ref_doc_id": "7d6b85a8-1c1c-4b4a-8c9f-9827ff561d18"}, "f33ded4d-b4d3-4685-93b3-bc7231ac85e8": {"doc_hash": "dd4c81bdf5cc720a0182dfa50b3ac880c6e0a70d9a2cba9d3a5b9796c4868082", "ref_doc_id": "7d6b85a8-1c1c-4b4a-8c9f-9827ff561d18"}, "ab91e265-e1cc-463a-9277-8eb7f9086794": {"doc_hash": "0d3f9201045f0061d543df3e2c4453e99a79cf1cd665792ff33bc53d65a1e106", "ref_doc_id": "7d6b85a8-1c1c-4b4a-8c9f-9827ff561d18"}, "045590d6-758e-4abb-86e8-8a3a02191a5f": {"doc_hash": "6beaa3488ad35ce5022a7874772fe2b5844b639a881ab5a34e405baea798fba7", "ref_doc_id": "7d6b85a8-1c1c-4b4a-8c9f-9827ff561d18"}, "c83c8670-2419-4572-9a81-a859ba0021db": {"doc_hash": "e2f401db2e44f023edd75719be93e605b80866e603808a8df952e545c791d2f2", "ref_doc_id": "7d6b85a8-1c1c-4b4a-8c9f-9827ff561d18"}, "b3b9bab2-09dd-48d1-aede-2769a88f6bac": {"doc_hash": "78fdc4e45023779516b9eb7d941fac75b1a6588b24e1d9212fc85bd6ab55a147", "ref_doc_id": "d543d099-ce2c-4224-b9a4-ad759f438763"}, "61fa9f26-9135-48ed-9695-1adab3f9f588": {"doc_hash": "b1d62c37ddfd8039acbbb609d9e1526795117d6ba5b95ac48dce207100f9b164", "ref_doc_id": "d543d099-ce2c-4224-b9a4-ad759f438763"}, "f300981f-69ac-4e0a-b4d7-aa69fe74db59": {"doc_hash": "3a26e7368cb4ff139b0f2464924c0d1503f7eb0947fe8cf82e7e6807139adc74", "ref_doc_id": "d543d099-ce2c-4224-b9a4-ad759f438763"}, "b242dc15-7e7c-4e06-afa3-0a6a68e209c3": {"doc_hash": "23917c953572944b3cff98f9af78278dbc3fde1366271a6f61adf40193f41afb", "ref_doc_id": "d543d099-ce2c-4224-b9a4-ad759f438763"}, "c010ac8f-8b4b-44ba-8a15-52e55ecbdf4c": {"doc_hash": "b796160b18a64d86ded7257bfdd73e4e4a7e2c6bbfa583f19bfae4d32c0f2b96", "ref_doc_id": "d543d099-ce2c-4224-b9a4-ad759f438763"}, "d2f8d11b-631b-4434-8ae4-8da32d3a3437": {"doc_hash": "23824154bc612058ad6f7c0706df44d0735e96262e54eaa9ab55d62511ef7146", "ref_doc_id": "d543d099-ce2c-4224-b9a4-ad759f438763"}, "13e77e0e-de3d-4c04-a7fd-9684826b4420": {"doc_hash": "fdd25a443797eb4aa56631d388462c24ca5f4bd7db993ede83743219e1e73ebf", "ref_doc_id": "3290c02d-c27c-44d5-b841-aa705c70db9d"}, "82aa3f52-44ff-4ff9-8e85-01f0aa4f07dc": {"doc_hash": "9c8495b56d1ce297277487f810f01f6aa85cdf15dd0930cb34fbdf5620e73ac6", "ref_doc_id": "3290c02d-c27c-44d5-b841-aa705c70db9d"}, "f8bec8e6-d4da-4ede-96a8-c56cbb71f22f": {"doc_hash": "6a5c81acdb0196981bf6c68715c9251a3407eedc0f222a038d9c7ce6026e8103", "ref_doc_id": "3290c02d-c27c-44d5-b841-aa705c70db9d"}, "e43d10e5-2e85-4811-bca0-5cbb0dbf0907": {"doc_hash": "85b96d433d7f07c2855943c3f03dcf276425e13143af2ba945dc086dfbe35bf4", "ref_doc_id": "3290c02d-c27c-44d5-b841-aa705c70db9d"}, "764a3f60-296f-4bbe-9f8d-611a57d7b14a": {"doc_hash": "faa4ec5c4ca62a39cd2e7e625b5a919bd02c6289102ea00ff5cf060dd073b892", "ref_doc_id": "3290c02d-c27c-44d5-b841-aa705c70db9d"}, "80e29ef9-09f7-4ad0-b123-c0f084375603": {"doc_hash": "adf66470caf63df2eab7de4e9a7b3b2ec90ad819946edab1bc0162a075b43bfb", "ref_doc_id": "f4ace53a-c604-4e78-a848-ab7cd073ebde"}, "2f3fed68-fbc2-400e-bda1-721f056a8189": {"doc_hash": "a07088d44322f14ac0910b074c6c9419c3008af0c40a13178a3f18b863c61e4c", "ref_doc_id": "f4ace53a-c604-4e78-a848-ab7cd073ebde"}, "43935afc-b148-4e05-a180-fc64c144d4d7": {"doc_hash": "da4b60a075a09a68edf148106ba274b269806f7055ac719b8a2bf4b608f846c2", "ref_doc_id": "f4ace53a-c604-4e78-a848-ab7cd073ebde"}, "00a8f7a7-a346-4dc4-b8aa-60cfeeeab5a2": {"doc_hash": "c7b1172c228f54ba6abea690102bc2edd2e309309d7860f12b71bad6243a0718", "ref_doc_id": "f4ace53a-c604-4e78-a848-ab7cd073ebde"}, "f2712e0a-b6d7-462c-be38-6574eaf31e7f": {"doc_hash": "eeb7bc68de0f66b1ee27a66586a65feafcc53c868c167b49f62019f99fbf0c5a", "ref_doc_id": "5b21b249-9b0f-48d8-b500-41fb636aaa28"}, "c61fc7bc-98e3-4656-aa93-f11f9245629b": {"doc_hash": "f5fc81f803cd477dd4f09bbe232bb8c3ff6a499171e089f1af64aeeae42ee649", "ref_doc_id": "5b21b249-9b0f-48d8-b500-41fb636aaa28"}, "011cff06-fd54-4e89-a84c-14a498dfefc7": {"doc_hash": "5a5adaa9f9eb48f34641b3af2cdb89d891e8eac77c4092a79fa7bef9aa9f6076", "ref_doc_id": "5b21b249-9b0f-48d8-b500-41fb636aaa28"}, "2ebd71ad-f2b3-4aec-9f3b-505e92f90d1c": {"doc_hash": "474e62277dbd423d819ea32cad4f364224dead6a3efcd4e99ebe4fd6753fcbce", "ref_doc_id": "5b21b249-9b0f-48d8-b500-41fb636aaa28"}, "826655f9-172c-46dc-99bf-1017a8305b63": {"doc_hash": "a703b0583c3a3e8dc4bebe3c09efcbb82cf8f1e9f1e9bc8b02da9deabfe782cd", "ref_doc_id": "5b21b249-9b0f-48d8-b500-41fb636aaa28"}, "3b47d874-dd1a-42d2-a582-34bad3d8b0c9": {"doc_hash": "e8305a66b6aba32cea625d96c0146f0e9b253d2f01e1595b4b23653bddf2d6ce", "ref_doc_id": "5b21b249-9b0f-48d8-b500-41fb636aaa28"}, "edd73257-0ab1-45ab-abc8-54d97a7edd0a": {"doc_hash": "41fb5363715b09682e8f7e309375712c2befce1b2790cc7c140a1d7d042644c5", "ref_doc_id": "b26ce5b7-5cd1-4c4f-ab00-7f0ec4074897"}, "d8e60937-b046-4729-83f5-35b2bb6e3498": {"doc_hash": "184e7dc8fbcce3e3a209af53417383f490d23a0694120a070295fb7af96aae2f", "ref_doc_id": "b26ce5b7-5cd1-4c4f-ab00-7f0ec4074897"}, "e40e0d0c-2842-4989-9e31-f6cbe20c35ab": {"doc_hash": "e4e94118d0faf899b4763ba56f9bc3dc8acb0139261fa52194a7885f1ac2151f", "ref_doc_id": "b26ce5b7-5cd1-4c4f-ab00-7f0ec4074897"}, "4bb4fea5-d8cc-4249-99cb-d5a06b1528ac": {"doc_hash": "2ee34e2d61fe9cde19a09aee48db9ac93e9a8c5c0664db67c66354726a06d9b1", "ref_doc_id": "b26ce5b7-5cd1-4c4f-ab00-7f0ec4074897"}, "60bde3cc-3c74-444d-808f-5d7a1bbcea2d": {"doc_hash": "28621a41b00d85d0db1d6e97e70171ef0639bed459c4a538ade311acbd4ed3ad", "ref_doc_id": "b26ce5b7-5cd1-4c4f-ab00-7f0ec4074897"}, "f961527e-ce85-43f9-a7fc-dbf858464d20": {"doc_hash": "7710aaccdb83f10739596307bbd8d1f5520721f78384f02a4f27711d2231d803", "ref_doc_id": "b26ce5b7-5cd1-4c4f-ab00-7f0ec4074897"}, "035a0bb6-ce4b-4b04-bf6e-bda7e200b914": {"doc_hash": "e61a0b877ff01c0a3ded5db8c19fb532c42d33db9f1b3fdb98bba3e3df80145d", "ref_doc_id": "8500b2d6-015e-436e-a558-aa4efc073e92"}, "eda5aba7-a3f1-4bf9-9dd3-eaa15059a4aa": {"doc_hash": "a54f2eab654e2e91065e4ad575d49c34ba4f8cedb8ff7442260587f81faf2044", "ref_doc_id": "8500b2d6-015e-436e-a558-aa4efc073e92"}, "ab216420-b362-48ea-8290-5b42c8350509": {"doc_hash": "49262ce4c7c49a1a1e93635325b09b5a6275c4c2709297bc241f4e88d567a2b6", "ref_doc_id": "8500b2d6-015e-436e-a558-aa4efc073e92"}, "e9a91b17-9249-41b8-8ec9-c8c477194fcb": {"doc_hash": "f00e68ea28e28e2cdb7303198ba59dee01846a87a071ae643efd48278c5fe116", "ref_doc_id": "8500b2d6-015e-436e-a558-aa4efc073e92"}, "c5b7cc46-849c-465a-8715-8d69c0f9334d": {"doc_hash": "a644b7925325bc94a6b0c91b65bb479118c8667e1838b2ba9d99eb7a67432d18", "ref_doc_id": "8500b2d6-015e-436e-a558-aa4efc073e92"}, "bd4d0ef9-9a64-4eca-9203-80315058db15": {"doc_hash": "3c3419932d293a628a6a2f48c8b8fbb6eab31381a242fdf65d04bb4bfddde364", "ref_doc_id": "8500b2d6-015e-436e-a558-aa4efc073e92"}, "f8b751bb-40f9-4123-b6fc-5575dc5f2383": {"doc_hash": "1bbf85a6280a445b8b5e3bf5c1904529d0be184df24047edc1725801ef329848", "ref_doc_id": "8500b2d6-015e-436e-a558-aa4efc073e92"}, "c7ae3fa8-e717-4a52-9b40-c15d32848a2b": {"doc_hash": "286c256e8b51d6ef7260935ec604419d15f0772ed059c55e83119c162de7233a", "ref_doc_id": "8500b2d6-015e-436e-a558-aa4efc073e92"}, "ee7c1408-a55f-44ee-ac96-0fb40bbda6b8": {"doc_hash": "7f667442b5ce6125dac97f622193515dd1bd2ff6ae418a0558441a489bafdb69", "ref_doc_id": "c69dbb5a-e5cd-4e5b-baec-971425360fe2"}, "7f8e09db-ab4a-4a4e-8ecf-b23c92a915c9": {"doc_hash": "8f679c98aec353d1566cb8ebd24a71adb7299ecc1fe9de77c735e60f5a34a50c", "ref_doc_id": "c69dbb5a-e5cd-4e5b-baec-971425360fe2"}, "dbaf40b4-9063-4dd0-8a24-da12f3dc98cd": {"doc_hash": "f3bedd53394428e3e657ad817bb73234f06880960d6be8ba09263985910bcc27", "ref_doc_id": "c69dbb5a-e5cd-4e5b-baec-971425360fe2"}, "faf2efec-9e29-4db3-8c8a-ec87e4f3548e": {"doc_hash": "2d23682d97f5fc5269aef7ba551836af9349aa24144002b273cfaed70066daa5", "ref_doc_id": "c69dbb5a-e5cd-4e5b-baec-971425360fe2"}, "4b8db73f-108b-441c-be29-3db8f6bc5d78": {"doc_hash": "cc9275ec106841a0f3201ce691e9ce48bffcb38097fa1cb41fb6a5f0692fbc30", "ref_doc_id": "c69dbb5a-e5cd-4e5b-baec-971425360fe2"}, "ba8f4e68-acd7-4a47-a776-80b310febb8f": {"doc_hash": "1e51c6f880b4f6cdc099ac1e55bb9f30eb6b328adc9ddf444f06a2c90de1f4eb", "ref_doc_id": "721e6fe2-2be3-4029-8b16-c9648c3e2621"}, "ef19d03d-3442-458c-adef-cb8a162d42bb": {"doc_hash": "1af89e7068bbd3ab2280f2b480dc3224b549213e5ab812dc2b207d3c53294901", "ref_doc_id": "721e6fe2-2be3-4029-8b16-c9648c3e2621"}, "14a82aaf-27b8-4cf4-b149-7e5d5eceb628": {"doc_hash": "23ad7311f098bef6763b40a8a9da8fa3fd625aa362900bb3afa32365204072bb", "ref_doc_id": "721e6fe2-2be3-4029-8b16-c9648c3e2621"}, "228042aa-b36d-4bfe-a3a7-e4c838cb1a8d": {"doc_hash": "8ec9692e1ced98ad827c1fb059457fd838cca7c7da8c0627bb09dcf726b028fe", "ref_doc_id": "721e6fe2-2be3-4029-8b16-c9648c3e2621"}, "33d77f1b-639c-42f9-8483-4026f5c09bef": {"doc_hash": "6b7c344e46f7343fa01fe820a04d8f5297ae165105197f1b083d64798f08c5d7", "ref_doc_id": "721e6fe2-2be3-4029-8b16-c9648c3e2621"}, "f5231109-65d5-43e0-8568-e77d8249a5ee": {"doc_hash": "0fef31a920157b4546829aa7e212da895c812ba0d3183927e74983f2cbb08f7b", "ref_doc_id": "721e6fe2-2be3-4029-8b16-c9648c3e2621"}, "4628dd81-f2b7-4da3-8457-109259410331": {"doc_hash": "2645d0af1c3bbadf123eef3e450d6f901ae1e5d8207353035e090317a2926bb2", "ref_doc_id": "721e6fe2-2be3-4029-8b16-c9648c3e2621"}, "caf9301f-cc5a-4f40-bfae-626ad1968ad7": {"doc_hash": "99f1f586af940173cfe6781f9be6ad4df1eb3382e70548b722f16fec2cd4c0b1", "ref_doc_id": "721e6fe2-2be3-4029-8b16-c9648c3e2621"}, "77d4247d-6f97-4b5e-af12-9949a43daab5": {"doc_hash": "1d559f3b7028d829f79f480a8db993c91148629e2b55b3613bad469e3f18290d", "ref_doc_id": "0c16360d-0190-4875-9f56-c31cfa1645d9"}, "1a36b08b-4d79-4031-8b3e-cb2cd47bf4e2": {"doc_hash": "4a08db69800892fd187de93bc0a5d9dc0ec3e50d21b76b397b699ba84ac604c6", "ref_doc_id": "0c16360d-0190-4875-9f56-c31cfa1645d9"}, "d4cc0b60-4559-4a49-b3e2-48aa039b2322": {"doc_hash": "a9f7e729eb71a11a557a729afa458eae4dd30e660a3486dfe9a914e9bbbab56e", "ref_doc_id": "0c16360d-0190-4875-9f56-c31cfa1645d9"}, "837a6005-5f03-4a63-a9d1-4e2092bbe90b": {"doc_hash": "44b4655aecdb1c853058d76abad75f06a29b5c397186871867d23a4a47a9a41f", "ref_doc_id": "0c16360d-0190-4875-9f56-c31cfa1645d9"}, "de05702f-74c6-4c8d-a3d6-2c6af282b163": {"doc_hash": "0698645e4e85df0b593529797f9c4b92d7353c4ee9d1b211128a460bb581debd", "ref_doc_id": "0c16360d-0190-4875-9f56-c31cfa1645d9"}, "384fb0aa-5011-4a5f-a17c-4e27b0d2d10e": {"doc_hash": "fedb3e813a417d3e49942112ef9f933155dde3dd7d22c5e966efbf7fe77aa053", "ref_doc_id": "9e4a0704-7ac7-4488-a38b-21ebb11fe736"}, "76dfbd6e-2f7c-4dad-ba67-b8f0541ebd08": {"doc_hash": "9fc347ed0f0720f0f9697ccec3ddc7da8d78a93a52dd67b74405fa320f87779a", "ref_doc_id": "9e4a0704-7ac7-4488-a38b-21ebb11fe736"}, "9a79589e-b750-4e08-9e6c-0d47148f7d7c": {"doc_hash": "b4ad8e93362d0535b481fafd2f35c2d80495dd79154f178cb47fd975b27eefa5", "ref_doc_id": "9e4a0704-7ac7-4488-a38b-21ebb11fe736"}, "5d153fee-4bcd-46eb-be08-e668159ce954": {"doc_hash": "b25c72c0770d3a745ee2df1b4283349884ccb7ef549da9a522f24c4c1de69f94", "ref_doc_id": "9e4a0704-7ac7-4488-a38b-21ebb11fe736"}, "b163aa80-b9e0-45d9-ac32-f3d3df8ff57e": {"doc_hash": "c0026d207c549cd92d80fa862c2498a216cf94807d64491e6614a7dcfddf2c8c", "ref_doc_id": "8f18d25c-ff85-49d7-a478-11d026ab472f"}, "d50f733c-49de-4759-8e76-2362aa141eca": {"doc_hash": "3e03fec6a52f958b897685045865c9f13c30ea1d22ea390b591fe9e25df7f0b9", "ref_doc_id": "8f18d25c-ff85-49d7-a478-11d026ab472f"}, "19022137-33ca-4bc4-9c25-84583e635017": {"doc_hash": "91437bbe0b922a5587814739b7a93aa6708472eaaa484db7e2789cd806ec3451", "ref_doc_id": "8f18d25c-ff85-49d7-a478-11d026ab472f"}, "1cbbd6e3-e7ce-47d5-a560-d15fca7ecc58": {"doc_hash": "098d898699ce3eabc62fb7e6ca223d07baf98ec392f4cc447b4abd452ae7cc90", "ref_doc_id": "8f18d25c-ff85-49d7-a478-11d026ab472f"}, "c81e1197-5c10-4242-9505-acf800451254": {"doc_hash": "95e564e82179d3613ea61dd07d9dc8ea8c647ce9020fd7344a687bd4b264681f", "ref_doc_id": "8f18d25c-ff85-49d7-a478-11d026ab472f"}, "6283fc3c-7275-4b4e-8d9b-50df3aebb288": {"doc_hash": "c1f258daf89ca0070248f51f0809cd3a059392a1a78d6280b6863b8c8a9c080c", "ref_doc_id": "8f18d25c-ff85-49d7-a478-11d026ab472f"}, "57e041d4-db81-42b9-b181-900c2666dba8": {"doc_hash": "e8cf7caa3ce3516a335ace2a7c7c0a688b32650f276cd2dacc2af1af228fa134", "ref_doc_id": "8f18d25c-ff85-49d7-a478-11d026ab472f"}, "f5ded79c-9fee-4f07-8065-576f28712bea": {"doc_hash": "fdfc4783d4172c9d5956dfc26ff42d1c830acde8b1f0aefbeb1f5031e686b296", "ref_doc_id": "8f18d25c-ff85-49d7-a478-11d026ab472f"}, "062d9cab-f298-462f-af4b-0c2cfb52a456": {"doc_hash": "31dbf21418eec94113cb6b19f6925b9c2e20f611b426f47004fa03566e45c0ad", "ref_doc_id": "8f18d25c-ff85-49d7-a478-11d026ab472f"}, "39e4ff77-a4f5-498e-a2fb-94f577ae7908": {"doc_hash": "83e8e42125e4c477446b67f8b03fe3a6b437d870c4b11109505de14fb0b823bc", "ref_doc_id": "b1879836-b599-41e1-af94-4bfe72a77860"}, "a4b15fbe-d158-4265-b5ef-7acf0309dbb3": {"doc_hash": "6d0c1f0fa5ef555549109e642e9a987e4eaf08ec2941df7e7ef9575b54dbc56b", "ref_doc_id": "b1879836-b599-41e1-af94-4bfe72a77860"}, "856e13fe-3def-4ed6-b440-d474562d6790": {"doc_hash": "2e5bb07417d7a1c0f9a9fec667832144d3ccf5f14121a46cb977eec81ed7066a", "ref_doc_id": "b1879836-b599-41e1-af94-4bfe72a77860"}, "3f8b5798-e6c2-4a6e-b2f0-a8c043396824": {"doc_hash": "2e32b17668687507375d9641ae8888ec8c1c824f6beec7543287b93444894d81", "ref_doc_id": "b1879836-b599-41e1-af94-4bfe72a77860"}, "8cfb97d7-4310-4249-9ead-f100246bb15f": {"doc_hash": "b900da4d37d0e8dd2c094026cd6322a09a24d509d8719818869308bf7c8b22b5", "ref_doc_id": "b1879836-b599-41e1-af94-4bfe72a77860"}, "7d0469e6-407b-40f2-8c73-5d72911d8f27": {"doc_hash": "a201bceaea49e010ce3e3ada816ccd2f31a79fbd86ce3f9707f9be12b5950a98", "ref_doc_id": "b1879836-b599-41e1-af94-4bfe72a77860"}, "ab357d46-6c15-4fb7-98af-1eec2922bd7f": {"doc_hash": "e088821ba90961cf4c053b97a8eec59e5e7d7ed0145abd7dd917cc2ce1e2240c", "ref_doc_id": "e4af41c9-853f-4975-8439-415e32a44762"}, "d268d7b3-399c-43ba-b818-05417e289f3c": {"doc_hash": "b13730e58f2872a15c64afe9f72661c2535fd2df2c368c633b0828a063d67660", "ref_doc_id": "e4af41c9-853f-4975-8439-415e32a44762"}, "81acb7c7-dcff-4da1-b3d0-259f2d86fa79": {"doc_hash": "b9c6dafee08311273588d181f3bb8b57af6154afbb095703e99cdb0e01a5564b", "ref_doc_id": "3c2e8b58-e9eb-4f7a-884f-41f0155d56ef"}, "fc8366b8-f6a1-4bd1-ade4-041f326ac4dc": {"doc_hash": "7172efb129356757dd5c7aaa5e79680c5b63afa51ea9cdfb5fd5204006bb38f9", "ref_doc_id": "3c2e8b58-e9eb-4f7a-884f-41f0155d56ef"}, "813b7a2f-e1b4-4784-96f3-1be0370c1b22": {"doc_hash": "29e6f47010061b9a69ff8ddf02fee954e777e87990df9d7eac17ea5f987f34f3", "ref_doc_id": "9d4c0667-0535-499e-afcc-b195f78cb549"}, "f00a11b4-bbda-48c0-a484-294f58a4e657": {"doc_hash": "5c4540d2a91fa3dc4ef4ce92a346d2f51d39e554ebc72f081e6d4be6a6e5a440", "ref_doc_id": "9d4c0667-0535-499e-afcc-b195f78cb549"}, "3b03f2dd-94df-43a3-9f25-751fd859ee56": {"doc_hash": "d7838bf6e499d86183cc627da2807bedc070a3877130d3a20d5217769b1f4d29", "ref_doc_id": "fc130a71-3273-495a-8425-cb428aafe536"}, "560eaa93-8cb6-4f32-b481-beb417607cf5": {"doc_hash": "de9959748126d2a010c2c18db36050f1c968cca1fe836d281300b8976908b140", "ref_doc_id": "fc130a71-3273-495a-8425-cb428aafe536"}, "48260b1b-0415-4f80-b0e2-30144e171771": {"doc_hash": "5aa51cff56be57baa4cb10c1c1cdacf09b63b5bff9d356db9227b35e33e353ba", "ref_doc_id": "9bc994d8-83ee-4424-aa7d-cab1758f2d15"}, "f8d21845-0325-4ffa-a0b5-0718e6079c93": {"doc_hash": "3a9ab2018d9efdb8823c3769fe77a9a66946a93a3bc0c8dd835e30647e09baf4", "ref_doc_id": "9bc994d8-83ee-4424-aa7d-cab1758f2d15"}, "7e4e6d40-6aa8-4450-845d-c01cb3d0879c": {"doc_hash": "087f635b21ff7594f883b7e019bad64e1ca1d37d10959f306b65e18768dc9537", "ref_doc_id": "bd8e03bc-e601-4841-824f-ed9aeec49953"}, "8ee342f5-c56f-40b6-8ab3-374f48f95ecc": {"doc_hash": "217e6b4b19a89b7c20815788fc428cd116acfcf20bca540432079de9f4979f13", "ref_doc_id": "bd8e03bc-e601-4841-824f-ed9aeec49953"}, "74e7fd9b-b060-48cf-80ce-3e8c7bd02324": {"doc_hash": "390a4309d4d50cb9ef7e07d7a8844070cd74628df663f796f1e5936ee9958153", "ref_doc_id": "4d5cdb46-0352-41e4-9e9c-957f124e0967"}, "b1239ece-1757-4579-9a39-e084b8c6f1c8": {"doc_hash": "c86ea884c7dafe91d43ae3944367ba7a92fb87647a46f74d5787bf7498718274", "ref_doc_id": "4d5cdb46-0352-41e4-9e9c-957f124e0967"}, "ae7bdbae-9d01-44e6-8f71-00f59971ea39": {"doc_hash": "0b532f10a66745b59483d59d092b62503179cd65d46c2b21799315ade3ff5410", "ref_doc_id": "2e1c4628-7cdd-4f08-8d86-072b77ae9761"}, "0d9a1aed-8f69-4212-80e7-44708a37f5e7": {"doc_hash": "1d53072bc20a297afec55ff39fe0121626e11660020eef0226cab0537193ef8e", "ref_doc_id": "2e1c4628-7cdd-4f08-8d86-072b77ae9761"}, "f64a2da0-63e9-404d-9e95-e8d462c4da03": {"doc_hash": "4b7953c1a33b54ce227322e1ffa98f4bddb8671d8f5af9a123509b1c1c426d63", "ref_doc_id": "6de21fe1-d7d1-47c2-be30-c83697669eb5"}, "5877d42f-f6c7-4b0c-a003-cc67dd788599": {"doc_hash": "5c94160995dc0468555255e891b2069b8f83210ceda577a133c5f8c75e3352b5", "ref_doc_id": "6de21fe1-d7d1-47c2-be30-c83697669eb5"}, "5eb87bf6-d65a-494f-9093-4eb198babbc3": {"doc_hash": "81e51deeb46c7f596fbd7220773cfa5206fcfb0cb511d350bfa751c9a177e04d", "ref_doc_id": "0de2524a-66fb-45c2-8a35-7ad0f2e66e36"}, "6672141f-5ff2-44e3-95d9-49dd3bfd4b97": {"doc_hash": "1770fe428768b7656b5f49068ca32ad8c5a31fa19be597cf100fc5ad43be5b73", "ref_doc_id": "0de2524a-66fb-45c2-8a35-7ad0f2e66e36"}, "517ceb00-a21c-43fd-a3e1-999e2a837bc8": {"doc_hash": "7c1f1267fcb112fff88743441d95a64d1a90c606163be178e9a523436c9a7c54", "ref_doc_id": "f9d2322a-f237-46b4-939a-b90f81d0e922"}, "32613ead-e644-4000-97ff-2bcad3244831": {"doc_hash": "7d253b9f0eae452661d78fe063159247e0fb13b97fe96af5507af94763eeb6e5", "ref_doc_id": "f9d2322a-f237-46b4-939a-b90f81d0e922"}, "41f6e3be-c7e8-483d-a345-4b16404411de": {"doc_hash": "723df848cffc6f2a1c2e88b1a8a302a0651ef07eb8f2aa776b45f1b7b82fd633", "ref_doc_id": "0b6c8e68-14a4-482b-a5a8-0e70a36ba52d"}, "f6bf074d-9de7-49e9-81be-78ff8f27d9b1": {"doc_hash": "1bdc619de56fe859600b6797b824846005b3614712b6f57aa4fc23e1ed20aac3", "ref_doc_id": "0b6c8e68-14a4-482b-a5a8-0e70a36ba52d"}, "e12b8d47-e893-40c2-af6f-cf52cf874035": {"doc_hash": "94dd88755af7df284a9baf69633a90eeb66fa1c98cf73ca61bf47ee7bf86ccaf", "ref_doc_id": "87f7dab9-8506-4d0c-9b70-4cbc0003cd75"}, "63c50146-9ae5-4fbf-bd1f-92f1f0bc0255": {"doc_hash": "aa808aa11f3f8ea3167abcbdd5d5b4ef7c4c1ccff0d42217dd0c874d34d50cea", "ref_doc_id": "87f7dab9-8506-4d0c-9b70-4cbc0003cd75"}, "117b0747-5d77-4e73-a31a-003d96ed1e86": {"doc_hash": "3331eb2e9de376bd019f697988de00aee747d7f51314e26edfcb48f88996e1f2", "ref_doc_id": "672b007c-4b9d-4463-a5d6-c4e1be707ef7"}, "89b8dfba-f448-47bf-9f08-711337807584": {"doc_hash": "0e8db2dc7df6f48c5a0ee2855630bb70215954205d023bfc5573cf245bc91165", "ref_doc_id": "672b007c-4b9d-4463-a5d6-c4e1be707ef7"}, "a9fada8e-463e-4063-8ce5-6f47fc12f666": {"doc_hash": "0c49270fff53a75436702a517c26a4558e93d6dbc7591ab9d139f5b5e75743e9", "ref_doc_id": "ca0fa447-93f1-41a7-b071-603606dd300e"}, "74c41a08-73c9-4d8e-8c3f-aa24daa056f7": {"doc_hash": "ae590a26baf3041d9a0adc6974856f733276247b209267448325b3748a4c6bf0", "ref_doc_id": "ca0fa447-93f1-41a7-b071-603606dd300e"}, "cbf994fe-1a71-4fea-bf84-67df6b4bab6e": {"doc_hash": "38e092916bc66c29ddc919918fa10b4570bc96cc32a898ab0026d504140b8430", "ref_doc_id": "2463fbe5-c01e-4300-b6ec-eca629caa29b"}, "14415354-caa9-443b-9f88-a0e545afff92": {"doc_hash": "e0265a62ff9d865824af917e710f277a3c9e6e446605ae199b1b617bf32c7e1f", "ref_doc_id": "2463fbe5-c01e-4300-b6ec-eca629caa29b"}, "4c21dca7-54f7-4706-a32a-350f74581f7e": {"doc_hash": "df13aa9016434e6003c7d97406f6183174a1b7eeb740d92837f2dd25b477c632", "ref_doc_id": "2463fbe5-c01e-4300-b6ec-eca629caa29b"}, "6f0b9627-227c-4eb5-92d1-4a57e4af6736": {"doc_hash": "8544dc949beb6f03d6b8aa1300ec37be12b7e4e62a97302c5aeffcb8f1c3a074", "ref_doc_id": "619d0736-2192-4da9-828a-f33ef9a047db"}, "bea362d2-325d-497a-97d2-019c0ffa3e61": {"doc_hash": "582d8920bde34c9914d665e629084c8a2324dbe721ac5b880a30156637330760", "ref_doc_id": "619d0736-2192-4da9-828a-f33ef9a047db"}, "13b2c56c-d4b4-4173-b755-8b9800d99c3d": {"doc_hash": "e6cfce1adf21357882dd73f52514f3b9cca9fc9e63e54b5bbc9b62e656bfcea7", "ref_doc_id": "619d0736-2192-4da9-828a-f33ef9a047db"}, "84388c23-78e7-4d73-9103-a443e9587e62": {"doc_hash": "0af21273726309b0bd66dab07087b2d92fe340084e862ed967ebdad337345328", "ref_doc_id": "619d0736-2192-4da9-828a-f33ef9a047db"}, "36c8ac35-d0df-4e32-888a-72bc36a252d2": {"doc_hash": "8138297f21fd97d93f848933c7cec8289ecb7c9358ffef0bbb52e3c74c0b8bff", "ref_doc_id": "a3e1e120-6732-4b4a-89ee-477130998e0e"}, "e244e139-e9a1-4e34-92b2-5835ada25da6": {"doc_hash": "65a1d54feff714ad767fd7186908f07f1a15df9d9fc4f9199a11093a0e6c722b", "ref_doc_id": "a3e1e120-6732-4b4a-89ee-477130998e0e"}, "bc9e8ddd-384f-4f71-b00b-acff631de8d2": {"doc_hash": "543060de817214355f744990f57ac20fe01963d895658cd73bd8c906491a73ea", "ref_doc_id": "a3e1e120-6732-4b4a-89ee-477130998e0e"}, "ade4f093-5318-40d7-9b81-2204d26fff8e": {"doc_hash": "b270217354ba21822ccd1bdd1f5f30a1c001697cdd7e48ad448a3a6a590ebb9a", "ref_doc_id": "a3e1e120-6732-4b4a-89ee-477130998e0e"}, "8e1b103b-1d79-485b-8bbe-3161ba6db37f": {"doc_hash": "7ec7bbbeae3a50b6d5b7e41c5fee182abb0d778ff4bbf8f6311b9b23e6d1fa12", "ref_doc_id": "a3e1e120-6732-4b4a-89ee-477130998e0e"}, "3760465e-0f22-4a23-9bef-025caa471241": {"doc_hash": "6bc5272bd6eb89a4abff207b1237b02ccdaa68eb469d6a2fdbc1c7156af9ece8", "ref_doc_id": "1dd47884-7047-44eb-8ed3-82f98752a400"}, "4bf4e7f0-65c0-4f05-a1ee-fd3e3a61809b": {"doc_hash": "b43cf58a04705f81a57445813a2619c4424de5b3b22223e0c361c50401b5cb08", "ref_doc_id": "1dd47884-7047-44eb-8ed3-82f98752a400"}, "1354b19a-c9b9-44fe-825b-246ecf80ea48": {"doc_hash": "18e7f737c5dcb40a9e703da8ff1f6e2f2bd1bd9825aacc31e7d39b312027ab76", "ref_doc_id": "1dd47884-7047-44eb-8ed3-82f98752a400"}, "00a98235-a28f-4d0d-8435-9396a1e73830": {"doc_hash": "9c0f619fee7c510fbac431d2aa90f7e3975d4faf6ee604af3f65e8df529ac559", "ref_doc_id": "1dd47884-7047-44eb-8ed3-82f98752a400"}, "b5b9b8c1-06ce-4c13-8251-cbc1faf88485": {"doc_hash": "e87f3e99d6ae50113ae9da4af76bf2d898a4070ca04f5fd00d4b13b3a5ca7edc", "ref_doc_id": "1dd47884-7047-44eb-8ed3-82f98752a400"}, "96e72751-1642-4479-9d54-b5a803ecee88": {"doc_hash": "ce4575eb24ac235736a67539b15792cb4228361d4b736d06607e5a3ddd3b6b36", "ref_doc_id": "56f63095-98b7-46bf-a7dc-790d001587c3"}, "2d8ae984-b0ca-4c47-b83c-b9c253d2a67f": {"doc_hash": "79ba961989126a75f4be747b2578a6fd4d1e8cc78377f4ef3aae0aa1a22335d8", "ref_doc_id": "56f63095-98b7-46bf-a7dc-790d001587c3"}, "c1eb1480-19ca-4c60-aced-2a66c976cc36": {"doc_hash": "19665b49b584e3fbb912c6a6ea21a6c93e8ac8675f6160930ffde8c131a9bf04", "ref_doc_id": "56f63095-98b7-46bf-a7dc-790d001587c3"}, "93fc5a11-ba77-4314-a9aa-69af38413333": {"doc_hash": "40b1ab85b3965bad6b7a4efe924b66c3a5b2ae92f33092c98ef1d7efe0ba84d3", "ref_doc_id": "56f63095-98b7-46bf-a7dc-790d001587c3"}, "c46ef647-4874-4fbc-be8f-80e3c1b4dcd5": {"doc_hash": "66b47a8caccc8fb81645ab707a21c1617797d58270c5f8b0dcf79d63fcf8ffa0", "ref_doc_id": "142f0092-e09f-411b-81fc-dd5b2fe76ffb"}, "eccc65c3-7dd9-4b2d-a4ef-4e495f0c90c9": {"doc_hash": "3143a62e3b9937ef5d08796020c1639e440cc1b3e62d06ade284705cc62a5a14", "ref_doc_id": "142f0092-e09f-411b-81fc-dd5b2fe76ffb"}, "fc8cafcd-884b-4f2c-868b-21b9b442fd04": {"doc_hash": "021c9882d98a9fe1f6527e8e3a00e260d2f3aa3772cf6a00712116f810704b0c", "ref_doc_id": "142f0092-e09f-411b-81fc-dd5b2fe76ffb"}, "f271ba5c-b90d-41b8-ba61-abd944580ccb": {"doc_hash": "b25f2599583729d2fa72026d3c4a204515eea6a632d29948776a3a542872c099", "ref_doc_id": "142f0092-e09f-411b-81fc-dd5b2fe76ffb"}, "6abd790d-8794-4c04-bcf0-f669a2529da4": {"doc_hash": "7c5d3613af88944cc27da24dd177df2b07b2e31ed468545c2f0ca0605a90617a", "ref_doc_id": "a6390232-e74a-4453-ac1c-f469dac6f577"}, "b42e822c-3234-4543-99f4-28f3def94b28": {"doc_hash": "0732b78ccc98ec4631cb333d824fec8392019913e1decb95436e32aa341ab944", "ref_doc_id": "a6390232-e74a-4453-ac1c-f469dac6f577"}, "6a2ab275-5181-4cc2-b554-cc30e92ebb1d": {"doc_hash": "9986ca430ee342fcbbd99a1af3db3ddab2f46fe11d556c9a8a663f4e64fc3150", "ref_doc_id": "a6390232-e74a-4453-ac1c-f469dac6f577"}, "f4110556-56cf-4a3b-854c-63133116eb05": {"doc_hash": "7e6dabbab3b0cc695a111ef6a9e56c5636a72027d37036d5fd662dceff805e3f", "ref_doc_id": "a6390232-e74a-4453-ac1c-f469dac6f577"}, "09cb40e9-ec9e-400f-9eef-1708cefdcd04": {"doc_hash": "850ddb33e0e729f6e3ac7204c485c2dbe27fc947cce59ea4fda17eb99be4b6e2", "ref_doc_id": "a6390232-e74a-4453-ac1c-f469dac6f577"}, "004441c6-45e6-4480-93d6-4f85f41c09d3": {"doc_hash": "8c9a03184f7b5c4e41a5a600a28042793359ce7f4f980577b47e4f98caa3f054", "ref_doc_id": "bb57f321-c60f-4e25-8444-a5cdbee87b27"}, "7ea36911-fd07-4f09-b748-1cb9ca04f495": {"doc_hash": "4fe2ca6d0ece35dde4a95d42f529f4f2f4e0691834d57b6bd2360621d8a9692d", "ref_doc_id": "bb57f321-c60f-4e25-8444-a5cdbee87b27"}, "20458bfb-7830-4b2e-a62f-8e15a800029c": {"doc_hash": "312033805cdf725f4b2885acb3ff6e2afac83fa5abb5117332f09a4752456ca4", "ref_doc_id": "bb57f321-c60f-4e25-8444-a5cdbee87b27"}, "13bf3d72-249e-4505-8136-7246fad2c783": {"doc_hash": "bf380eb278353fb25605ac29f65d35beaf7c9428ff01bd33ad52af6048b63161", "ref_doc_id": "70efd5d5-bcda-4698-a7e4-c0ba91bf514a"}, "f7527283-444c-483e-857d-e02388c28f30": {"doc_hash": "4faf3793a591690022d16bf82dcac57b9bc76ff4c9edf32d6de430a321b0de02", "ref_doc_id": "70efd5d5-bcda-4698-a7e4-c0ba91bf514a"}, "4b37021e-9968-4f85-b8b9-3cb8e730ad15": {"doc_hash": "05aa0d513f045c09501bc1c6acc1ada1f5cccde7c0bb58eb9b669d4ba4b4ea7b", "ref_doc_id": "70efd5d5-bcda-4698-a7e4-c0ba91bf514a"}, "36bb2f32-dc7b-4109-94a8-ec50e301c6a7": {"doc_hash": "9bcaafea70f31abe07fed514eb4cca9e371a25e05ad6126cf9fe9a0d1cace3ca", "ref_doc_id": "70efd5d5-bcda-4698-a7e4-c0ba91bf514a"}, "008368f5-ed54-4b4a-8e7e-f808ace630b7": {"doc_hash": "a7a87586ce64166ee85aa5f17a7b9258a209cef9776eefce570115032adee3c7", "ref_doc_id": "70efd5d5-bcda-4698-a7e4-c0ba91bf514a"}, "e2dd3028-2df8-4f0c-9631-86451db2c111": {"doc_hash": "39e0c15a3c93cad2fa6be3fc34e71dce10ef658d0634c553da6ee6473132aa19", "ref_doc_id": "7ee672fa-2873-49ed-bc6c-25464d3d468c"}, "67058179-72da-4f35-adfc-ccbaa8300641": {"doc_hash": "29fd4686272b7bea7e8ced1e9a934ad32d8a19a24f3cc5ceec2e4d6899a654e7", "ref_doc_id": "7ee672fa-2873-49ed-bc6c-25464d3d468c"}, "a21bd965-501e-42d3-b74b-c49995ad5f4f": {"doc_hash": "17862337f3045e577fccae9bae95e0979a5a496e8a26f3b6cba1a5b2cc3eaa6a", "ref_doc_id": "7ee672fa-2873-49ed-bc6c-25464d3d468c"}, "8054dd11-78a3-4777-b1ea-55132e226d1a": {"doc_hash": "b3a11c8c6cbd5451f280fd44be3f990f14436ad513bd6a4701484fc73d672043", "ref_doc_id": "7ee672fa-2873-49ed-bc6c-25464d3d468c"}, "89e86816-c6c5-42d5-bc8b-ab254044ee5a": {"doc_hash": "fa9362cdc2271a6e68e5e077e2d1fa84df56b0224af487a33287547ad92b723d", "ref_doc_id": "7ee672fa-2873-49ed-bc6c-25464d3d468c"}, "ceab7ae5-1fbe-4481-8356-475bf9774c29": {"doc_hash": "99dbeb532efa045df545242e5e1c9e39fb73f7cf00a35d0041636d93704903a7", "ref_doc_id": "7ee672fa-2873-49ed-bc6c-25464d3d468c"}, "cbfd04f2-7520-418b-bd69-a2200a9c66ed": {"doc_hash": "8e2cefc001cc15dce78e78c2a04e4a67982849f9a4b37a98a633e9cc7021a463", "ref_doc_id": "29d3becc-38ca-47cb-9332-674e4b7909a6"}, "ec2fabc6-1aaf-4704-85b0-f81e974ffb56": {"doc_hash": "bebf9c8a405f77ababb75fb0b31d69e96340d6d4ce837eda990cd2029dd79b45", "ref_doc_id": "29d3becc-38ca-47cb-9332-674e4b7909a6"}, "9ec2d462-902a-4147-98f1-ce3d1c978d93": {"doc_hash": "5430ad2917d6e4349804f6962b2792e8598333b93f5bf77d5eacdc3b4026ae6c", "ref_doc_id": "29d3becc-38ca-47cb-9332-674e4b7909a6"}, "68b2b3e6-15b0-4bfc-b376-00288ff26c68": {"doc_hash": "7b9c3bb7b19c99309614af9b5130cc88bbd97e2e92544909217178ae4a817137", "ref_doc_id": "29d3becc-38ca-47cb-9332-674e4b7909a6"}, "f9877c88-0e08-4067-b0ee-40dac187596f": {"doc_hash": "fb4873731585981ccb15e2af00b1c7f621baecf64cfd561157f788104454fd9c", "ref_doc_id": "29d3becc-38ca-47cb-9332-674e4b7909a6"}, "8a448c50-08b1-47f8-853b-919396983891": {"doc_hash": "4193f4744201a0499d6bed904b345b1213475f8f2acab1f748ccc55fa4fa355d", "ref_doc_id": "f3ccaf27-46f6-4aaa-b10d-195f2156f4b6"}, "96dc6fbf-2f14-4e07-ac4c-0f3385813d2b": {"doc_hash": "7409ba8a8ce9e2c430770b45f2be1640f1e68830657d69f8bbd4a9ab86fb72c1", "ref_doc_id": "f3ccaf27-46f6-4aaa-b10d-195f2156f4b6"}, "15062f1b-c7f6-45d4-8ae4-7fedc8eda488": {"doc_hash": "3e8fc50dc15c8f73919dc1191fb05765ef644b42ba6836c37457bd9f639a14be", "ref_doc_id": "f3ccaf27-46f6-4aaa-b10d-195f2156f4b6"}, "4c5cb8e7-665e-410f-920d-33413ad7c183": {"doc_hash": "2d6d1174409c645035140350372181fef32e1fc2e5b6e3e6a4f946b05f6057c9", "ref_doc_id": "f3ccaf27-46f6-4aaa-b10d-195f2156f4b6"}, "f9034dee-5640-4342-9a2d-c72ad0780d4c": {"doc_hash": "0c38ace2cff0650460187b948e01e9ed8076f663daa1a15228329535283081a0", "ref_doc_id": "f3ccaf27-46f6-4aaa-b10d-195f2156f4b6"}, "6c789761-7721-438b-83f6-32580146d47a": {"doc_hash": "156d3c031ec6488d878189fec780c8c6c7f0855d247fac0397ca2e0cc8b0e2ff", "ref_doc_id": "fb80ecac-79cd-47f7-a382-727a1c8a533e"}, "5243f89c-b9ab-442b-b607-6b8ea6277c0f": {"doc_hash": "41bbd130de97741fe94b0536e2c8bc13fb9e6a9a19e78e92377803b76c8a53c3", "ref_doc_id": "fb80ecac-79cd-47f7-a382-727a1c8a533e"}, "ee5db39c-b21e-4726-9736-122a98b25b66": {"doc_hash": "e49490c6bcd37155ac820840348febe5bf31f68b9921276910b99d9ad1bc741e", "ref_doc_id": "fb80ecac-79cd-47f7-a382-727a1c8a533e"}, "2b8a9431-524f-4944-8ab6-cc5042aed4e8": {"doc_hash": "5b6b07c565d0f75b40d29815bc99c83ed509e5165065a0f956de4860f71b66c5", "ref_doc_id": "fb80ecac-79cd-47f7-a382-727a1c8a533e"}, "a42737a5-ac14-4639-b0a9-050b2bc5b0c2": {"doc_hash": "a85f3d193333fac3bbc5de3ef46322e1d860212db3481334dddfe079ef9d7ec0", "ref_doc_id": "74571bde-6fcb-4af9-ab8a-35c53b325152"}, "b8e3088d-2326-4b7c-b6e7-6027432a2844": {"doc_hash": "498bb4fc7539838520c9a7c90a38a5dee2bfaf40da036313272a51177ebdd4d3", "ref_doc_id": "74571bde-6fcb-4af9-ab8a-35c53b325152"}, "b7294d1e-6da7-4f44-ad24-01f6c545c93f": {"doc_hash": "5521ad6ac58dbb10630a6a42a3a0bc325642f115c976b9ad06a880ac9733dcf5", "ref_doc_id": "74571bde-6fcb-4af9-ab8a-35c53b325152"}, "d5e7c57b-e43f-4036-a5db-b580d97735d2": {"doc_hash": "ae8b4959804b5fb46900dbf68d2c4cfa07c37e38333b58392c00e0bdddfc1a14", "ref_doc_id": "74571bde-6fcb-4af9-ab8a-35c53b325152"}, "fa64ff21-9041-4fa6-bf55-b326d25811e8": {"doc_hash": "c6774a8f38b5c6e1febf4370e2054e5884c1fccadf935ccbf51051af6535b186", "ref_doc_id": "74571bde-6fcb-4af9-ab8a-35c53b325152"}, "29fc01a6-e599-43ea-bcc1-2787e7698c86": {"doc_hash": "4f5ad3bbbe332963d4a269072a738a8323c8f22f04e1c04538ef070c1d363f9c", "ref_doc_id": "74571bde-6fcb-4af9-ab8a-35c53b325152"}, "34f3377d-b7e2-4dd9-917b-b5cd9adcfcb7": {"doc_hash": "09157afd1a3f5fbf1b308ff680aea36d887805195e44f86b4efdcf29ff93d707", "ref_doc_id": "74571bde-6fcb-4af9-ab8a-35c53b325152"}, "572359a5-bce2-4e26-9bdb-9369ef6e0c84": {"doc_hash": "bff8b55b39c96448d8341c8beb438fa19fb674bdb6f6f93886aacfb8277900e4", "ref_doc_id": "1deea7fa-6cef-41a1-96fe-bcf0ad499d43"}, "93b1b873-30ec-4f04-b951-d4dafe32d0c2": {"doc_hash": "6eea4aa343c0c5169579f08a755873fe349c7109f019a2449b207395d8812a15", "ref_doc_id": "1deea7fa-6cef-41a1-96fe-bcf0ad499d43"}, "96066384-e0d4-45fc-a5d7-3ff5958ec1fb": {"doc_hash": "fae9e8e3df301095aef41c725ee9537a68b0653cae7813e5432488c6324f889c", "ref_doc_id": "1deea7fa-6cef-41a1-96fe-bcf0ad499d43"}, "4482903a-bf89-422e-be1f-cffd4bf551fe": {"doc_hash": "19826f3fd7f2b0765cafa57d03c379457d2c4c76332081361beb6e5bc0cda9b2", "ref_doc_id": "1deea7fa-6cef-41a1-96fe-bcf0ad499d43"}, "fd9baaf3-ca64-4a04-a18d-2e4d634048a2": {"doc_hash": "0a019b8d2b96ee22634406129e35a3f740485245a372f7ed8b170df6a3535619", "ref_doc_id": "1deea7fa-6cef-41a1-96fe-bcf0ad499d43"}, "c5fc4b53-c961-43ee-93a3-a051764fb0c3": {"doc_hash": "0d32f8ac05493f5f7943d880735e10c5b1a4f2b763788497c3c003a2e90a2288", "ref_doc_id": "5acd044b-9d8d-47f7-b88e-cf4ea4514f96"}, "c7751fc1-42be-4936-81e6-4344d4449a51": {"doc_hash": "4b7e18b4b03ac594e2eafae429753d61997e4817f582ac75a5b3eeffb994084b", "ref_doc_id": "5acd044b-9d8d-47f7-b88e-cf4ea4514f96"}, "0a646399-e0de-4836-b25d-04ae232449c3": {"doc_hash": "2721aefb2ea148d524ebdb5c331b8924285d38999980af3676b4b89d8efe57c8", "ref_doc_id": "5acd044b-9d8d-47f7-b88e-cf4ea4514f96"}, "07a0d979-1c2f-46c4-a0a7-91de295cfd65": {"doc_hash": "f9668cceb3f9fce780248018d7cdd7b012af30c4d2d7b9f683c43c04f33c57bc", "ref_doc_id": "5acd044b-9d8d-47f7-b88e-cf4ea4514f96"}, "e85dffe1-a000-42dd-99a7-4c5eabf7ef14": {"doc_hash": "e1abf177dbf75c97fdb4440840c71e9f92470c14a3f8a96b6c0e666da7b02566", "ref_doc_id": "5acd044b-9d8d-47f7-b88e-cf4ea4514f96"}, "e9e11adf-a9e5-42a1-9eed-7b53e2289646": {"doc_hash": "6943be33afa823bcdb89b6ece2a9e82e766340743c3e9c7686f819ac5943dd17", "ref_doc_id": "5acd044b-9d8d-47f7-b88e-cf4ea4514f96"}, "67141005-9c62-489d-9374-bff1fa9da7ae": {"doc_hash": "27f98466bd354bb37217a806be518c5fa3c0835459e61ae2dca8ea5e29cde52a", "ref_doc_id": "1a293364-9a65-4363-a78e-033057f53944"}, "59b16652-6ade-45b3-9ad5-6bbaebd78823": {"doc_hash": "ba4a4a70a71e5842bdc4ca8a68ec95353e37393b06069320a95c570f1110a583", "ref_doc_id": "1a293364-9a65-4363-a78e-033057f53944"}, "9e946097-f0e9-444f-8e38-8402b039005c": {"doc_hash": "5a935b4368920ba63df09df4d1d0bd314b4c71e2c9766f5bf6d06f861d1bf37e", "ref_doc_id": "1a293364-9a65-4363-a78e-033057f53944"}, "894b0c9d-4aac-46ea-9b88-5450d1f33057": {"doc_hash": "3ef265b7016febbd86c5def6707ceaa709c9a70990ea8224a04f129bce3b3c6c", "ref_doc_id": "1a293364-9a65-4363-a78e-033057f53944"}, "1d88f0c5-ed50-41e1-99f4-b6bc66e49c64": {"doc_hash": "bd5e11acb5888d0d5cc8a108e7a9f3b109a362022de70d4a77d3463423cd10b4", "ref_doc_id": "1a293364-9a65-4363-a78e-033057f53944"}, "cb0ecf56-27a0-45f6-b6ea-9169066c6cb8": {"doc_hash": "3c96ebfbb84421a250d93b4a4f92c417d6ca697300c8dfe6b704c9c1522b1625", "ref_doc_id": "b0d761ef-9bec-46bf-a112-261d0f389c3a"}, "39a7d3e7-ab49-4d23-8cef-4bd4f8a5b3b2": {"doc_hash": "aa09ffc529f2d3ed161824e4e1331a53b4a4597965f66765dc217785eb67341f", "ref_doc_id": "b0d761ef-9bec-46bf-a112-261d0f389c3a"}, "e41dabeb-3e80-45bf-b38f-785a14c25b34": {"doc_hash": "d3a097eb66004504bf3902493c427dabbb81932ba1972cc61d7cee6f2fc89f22", "ref_doc_id": "b0d761ef-9bec-46bf-a112-261d0f389c3a"}, "f331d817-57db-4bd2-aedb-d960b44cf61e": {"doc_hash": "3de0938c1d709a1d287650a43f1601a613127713a6c049d566204f6b8907d47d", "ref_doc_id": "b0d761ef-9bec-46bf-a112-261d0f389c3a"}, "ce244400-592a-41ac-9992-aa96623a8a49": {"doc_hash": "7464148b739e3fb04a7b221ac723b04f23265d46fc625d6559739b5cf20c196f", "ref_doc_id": "b0d761ef-9bec-46bf-a112-261d0f389c3a"}, "69ee04f9-a2b5-4f68-a89f-5e0a468ed6bb": {"doc_hash": "355c4d033afd70d815cf458314f2d975fd4dfa78a59481fe17e3b91bd2ad39df", "ref_doc_id": "d64f1393-8ea3-4495-b60d-300797f050d1"}, "4fd74f11-08d6-41bd-a62d-3fe64bbaee4a": {"doc_hash": "43993cd17b3fdef9ac91d90a7b16ead6d23d00ca5bce40c63b400fd05a71444f", "ref_doc_id": "d64f1393-8ea3-4495-b60d-300797f050d1"}, "3ee06420-3e90-4e62-bdc6-a81fd84a1271": {"doc_hash": "a05214ed12213e7f306c2beacfea066665968eb8036cfc498d30eeabc43b70a9", "ref_doc_id": "d64f1393-8ea3-4495-b60d-300797f050d1"}, "0b71a2c3-b090-4c63-9246-2c508d56a785": {"doc_hash": "fb85dd74236bf749c9d90b8acbf4fe018886e380c1fc8b28a9ddc1266e90eb0e", "ref_doc_id": "d64f1393-8ea3-4495-b60d-300797f050d1"}, "99ac42b4-e3fc-40fc-9cc9-c28519b61054": {"doc_hash": "1286b76484809c8cc50b4663f5e6ae3e35aa12c14c3f871b6fa63bb2babfcca9", "ref_doc_id": "d64f1393-8ea3-4495-b60d-300797f050d1"}, "c87cd0d2-0b34-4099-88fd-ef65e4c7708a": {"doc_hash": "512d26260e67554603c361325b9ab4472a691c8841518ac812d457fda66f7c7d", "ref_doc_id": "d5c086c9-cde5-47c6-b8d9-3bd77a6a1b8c"}}}