Daniel0315/IDCite-ContextEnhancedRGCN

IDCite Context-Enhanced R-GCN

This repository provides a trained Context-Enhanced Relational Graph Convolutional Network (R-GCN) model for scholarly citation recommendation and citation behavior analysis.

The model was trained on a 500K citation-event subset derived from IDCite, a scholarly knowledge graph containing citation events, citation contexts, citation intents, fields, papers, venues, authors, affiliations, and geographic metadata.

Model Summary

The model integrates:

• Citation-context embeddings
• Citation-event nodes
• Scholarly KG structure
• R-GCN message passing
• Citation target ranking over candidate seed papers

The released checkpoint corresponds to the no-leakage setting where the target HAS_CITED_PAPER edges were removed from the message-passing graph during link reconstruction.

Files

File	Description
model.pt	Trained PyTorch model checkpoint
node_embeddings.pt	Final learned node embeddings
config.json	Model and experiment configuration
node2idx.json	Mapping from KG node ID to integer index
edge_type2id.json	Mapping from edge type to relation ID
model.py	Model definition
inference.py	Helper functions for loading and scoring
requirements.txt	Python dependencies

Model Architecture

The model represents each citation event as:

citation-event node embedding + projected citation-context embedding