Datasets:

Kushala

/

aionns

like 0

AION User Guide (Release 2.8.5) Copyright Notice The information contained in this document is the property of HCL Technologies Limited. Except as specifically authorized in writing by HCL Technologies Limited, the holder of this document shall: (1) keep all information contained herein confidential and shall protect same in whole or in part from disclosure and dissemination to all third parties and, (2) use same for operating and maintenance purposes only. ©Copyright 2018-2023 HCL Technologies Ltd – All rights reserved. Introduction AION (Artificial Intelligence ON) is an AI life cycle management platform for applying machine learning to real-world problems. AION encompasses the complete pipeline from raw data ingestion to a deployable machine learning model with a less-code/no-code approach. It includes the following sub-processes: AutoML- Enable business user-generated ML models with no code approach MLOPs - Model Versioning, Deployment & Drift Monitoring MLaC – Machine Learning as Code for automatic pipeline code generation Explainability- Higher accuracy predictions with supported explanations & model confidence In AION, all these processes are automated. When you give input data and the use case that needs to be solved, AION runs through all the stages of data analytics cycle and generates the best deployment model. AION Engines AION has a multi-stage pipeline as depicted below in the figure. A brief explanation of each stage is given below: INGESTOR- Data Ingestion: In this stage, the dataset is uploaded in AION GUI from disparate sources. EXPLORER – Exploratory Data Analysis: This shows the details about the nature of data, the relation between features, model statistics, and performance information to derive descriptive insights. TRANSFORMER- Data Processing: In this stage data-cleaning, data preparation, and outlier detectionares did automatically to improve the quality of data for better model accuracy.

SELECTOR- Feature Selection: The statistical analysis takes place to identify the relevant features for model training and remove the unimportant features based on correlation & importance. LEARNER- Model Training Hyper Parameter Tuning: This stage trains configured models and selects the best parameters based on hyperparameter tuning (using which the models are trained). A broad spectrum of algorithms is supported. PREDICTOR- Inference service: ML Model serving & inference services. OBSERVER – Model Monitoring: It monitors the model for input & output drift of data or predictions. EXPLAINER- Explainable AI: It explains the model & Uncertainty quantification of the prediction. CONVERTOR- Model Conversion for Edge Device: The generated models can be converted to any desired format that is suitable for embedded, cloud or edge deployments. TESTER- Model Testing: Once the models are generated, it can be tested using different testing methodologies. CODER- Machine Learning as Code: It creates Python code automatically for ML pipeline components. Launching AION AION Explorer is the application used to launch the AION GUI. Followings steps are followed to lunch the GUI. Double click on Explorer icon on desktop to launches the AION GUI. AION explorer shell will pop up loading all the requirements. After a while GUI landing page will appear in a browser as shown in the figure below. Users need to reload in case the GUI main page does not appear. Note: Access to Hyperscale like AWS, GCP, Azure and data sources like graviton and NiFi will not be possible on HCL network. To enable access user needs to raise service exchange ticket and raise access to this source. Compatibility with previous version:

Compatibility with previous version: Note: Note: if a previous version of AION is already installed, re-training of use cases of that version is possible in the latest version, and it is backward compatible. But, use cases created in the latest version may not run the previous version, as forward compatibility is not supported in all cases Model Management Create Usecase To create a new usecase, follow the steps given below: In the left panel, click Home. Click Add Use Case tab, enter the Use case name or keep it default and its Description. Note: Use case name does not support any special characters or space. Click Submit. As a result, the usecase is added in the USE CASE LIST. All the usecases listed in the USE CASE LIST can be viewed as shown in the figure below. Compute Infrastructure Users can select following computer infrastructure for training: Local AWS EC2 (AION) AWS EC2 (LLaMA-7B) GCP GCE (LLaMA-13B) Click on the Compute Infrastructure tab as shown below. AWS EC2 (AION): User can use the Amazon AWS EC2 infrastructure with default settings or can configure the details from settings tab as shown below. AWS EC2 (LLaMA-7B) and GCP GCE (LLaMA-13B) for LLM fine tuning. View Existing Models To view the existing models, follow the steps given below: In the left panel, click Home. The existing use cases and their training status can be seen. Trained models can be lunched for Retaining, Prediction, Trusted AI parameters and Drift Analysis. For multiple versions of use case, click on use case name to view different model versions that are trained by Model Retraining. Eg: By clicking on AI00053_lanquos, different versions under that Use Case can be viewed as shown in the figure below. Publish Model Trained model can be published in the form of following packages: MLaC Federated Learning Python Package Homomorphic Encryption Model Container Retrained Model

Model Container Retrained Model MLaC: Machine Learning as Code generates code automatically based on ML Operations performed during various ML Lifecycle. Using MLaC, expert data scientists can have better control over experimentation, optimization & deployment of ML Models. MLaC has four unique components. Generate Code: Generates the code for each pipeline component of the selected use case i.e., data transformations, feature engineering, model training and prediction. This code can further be used to retrain the model on the latest data and can also be used to create a container for each component. The user will find the code in the local system through path: C:\Users\yashaswini.ragi\AppData\Local\HCLT\AION\target\AION_375_1\publish\MLaC Generate Container (Local): Pre-requisite: Local Machine should have docker installed. Create docker images for each of the components by running the docker file present in each component. Once the docker images are ready to consume, create the container from the images and run that container sequentially as mentioned below: Model Monitoring > Data Ingestion > Data Transformation > Feature Engineering > Model Training > Model Registry > Model Serving User can see the status of each step through the shell as shown below. Exec Pipeline (Local): Another way to run this component without using docker is to run the command run_pipeline.py in aion shell. The command will execute each component one by one locally in the user system. Exec Kubeflow pipeline: Using Kubeflow user can run all the components as a single pipeline, which is managed by Kubeflow internally. GitHub Upload: Upload the Generated Code to the GitHub repository. MLaC service can be used for prediction through PostMan following the steps given below- Open PostMan desktop app. Configure a POST request with the predict API of the form http://<localhost:port>//AION//<usecasename>/predict Eg: http://127.0.0.1:8094//AION/AI0192/predict. Select authorization as Basic Authorization

Select authorization as Basic Authorization Username: USERNAME Password: PASSWORD Put the data in the Body as a raw json format. Click send to perform prediction. Note: For more information users can follow the user guide for MLaC by clicking on the icon on the top right corner of the pop-up window. Federated Learning (Beta): Federated learning is a way to train AI models without exposing the data, offering a way to unlock information to feed new AI applications. Federated learning (or collaborative learning) is an approach used to train a decentralized machine learning model across multiple partitions. The federated learning participants can be edge devices, data centers, smartphones, medical wearables, IoT devices, vehicles, etc. They collaboratively train a shared model while keeping the training data locally without exchanging it. Each participant downloads the model from the server. They train it on their private data, then summarize and encrypt the model’s new configuration. The model updates are sent back to the cloud, decrypted, averaged, and integrated into the centralized model. Then, this centralized model is once again shared with participants for the next round of federated learning. Iteration after iteration, the collaborative training continues until the model is fully trained. In AION, Federated Learning is supported for the following algorithms: Logistic Regression, Linear Regression, and Simple Neural Network (SNN). Python Package: Python Package creates a WHL File of a trained model so that the user can further use that file in testing and prediction. If the user wants to consume the trained model for prediction in some system where AION is not there, then either the user can install a python package or use the docker images. Python package can be consumed by following the steps: Click on the python package icon from the packages tab and as a result whl file will get download. Open the command prompt and type the command: python -m pip install whl_file_path and press Enter.

Model container: The model container creates a docker image of the trained model so that user can use that image anywhere, pull the image, and perform testing. If the user wants to consume the trained model for prediction purpose in some system where AION is not there, then either the user can install it as python package or use the docker images. There are two cased to create a docker container: When the Docker is installed in the machine user can click on the docker and it gets created. In case Docker is not installed, running aion will throw an error. For that user can copy the mentioned path file and put in other machine where the docker should be installed and can build the image there also using docker build command. The model container also supports CORS (Cross-Origin Resource Sharing) protocol to mitigate the default security policy, Same-Origin Policy (SOP) used by the browsers to protect resources. The same-origin policy is a critical security mechanism that restricts a documents or script loaded by one origin can interact with a resource from another origin. Command to enable CORS: python aion_service.py ip 0.0.0.0 -p 80 -cors “http://localhost:5000” Homomorphic Encryption: Homomorphic Encryption (HE) combines data encryption with privacy protection. A homomorphic cryptosystem is like other forms of public encryption in that it uses a public key to encrypt data and allows only the individual with the matching private key to access its unencrypted data. However, what sets it apart from other forms of encryption is that it uses an algebraic system to allow a variety of computations on the encrypted data. AION Supports only the XGBoost algorithm. To publish the trained model, go to the Packages tab and click on the icons of above-mentioned packages accordingly. Users can select any model or any version of that model for publishing so that it can be installed on the system for faster inference.

Eg: If an user clicks on Python Package icon from the Packages tab, a pop-up window will appear to confirm the model download. Click Ok and then the model gets downloaded in the form of a WHL file which can be used for further testing and prediction. Model Re-training  After the Model got trained go to the Home. From the Train column click on the Model Re-training icon to re-train the model withing Re-Configuring. For Re-Configuring and training click on the icon. Click On EDA if required or click on Next. Set the Basic and Advance configuration if required. Click Train Model tab to train the model. Model will get Re-trained successfully and user can see the results as shown below. System Setting To configure and change system settings, In the left panel, click the Settings icon. User can view the settings and configure them if required. Configurations for the following integration entities are available: Configuration Settings: Apache Kafka Configuration: Kafka cluster by default will be running in the local host. Kafka uses key-value pairs in the property file format for configuration. The default values of Kafka cluster ip and kafka cluster port are localhost and 6000 respectively. The value of Kafka Topic can be supplied by the user as per the Kafka configuration. Open AI Configuration: For open AI related task Open API key, Open API Base, Open API Type, and Open API Version need to be given by the user. Cloud Storage Settings: AWS S3 Bucket GCP Bucket Azure Storage Credentials need to be given for accessing AWS, GCP and Azure Storage by clinking on add bucket and data hence data can be uploaded from the cloud in GUI. Cloud Infrastructure Settings: The user can configure the details of AWS EC2 compute infrastructure to consume AWS EC2 infrastructure. Model Training The Model Training tab is used to train the model based on the data input file. Data Upload Data upload tab is used to load the dataset in the GUI to train the model and get insight of it.

While uploading the dataset, data size and ram size need to be checked in two cases: If file size > ram size, file will not be uploaded for further process. File size should not be more than 80% of ram size. If the file size > 50% of ram size than the alert will be shown and proceed for the next step. Structured Data To upload the data file, follow the steps given below: Create use case. Click the Data Upload tab. File like csv, tar and zip file can be uploaded from local, URL, NiFi, GCP Storage, Azure Storage, AWS S3 Bucket, and Graviton. Credentials might be required to ingest data form all the cloud storages. As we Submit, the file will be uploaded successfully. Graviton: It enables implementation of the modern data platform using polyglot data technologies, open-source data frameworks and cloud-native services. It facilitates the development and delivery of optimal data products and efficient management of data product lifecycle from development to production. Unstructured Data This option is used to upload a document. To upload a file, follow the steps given below: Click on the Data Upload tab. Click Unstructure data. Select the type of document you want to upload (txt, log, pdf). Place the files in a folder and specify the path in the Folder Path field. Click Submit. As a result, the files in the folder will be uploaded successfully. Custom Python Script The python script needs to be created by the user, which can be converted into a data file. To upload a python script, follow the steps given below: Click the Data Upload tab. Click Custom Python Script. Specify the python script path. Click Submit. As a result, the file will be uploaded successfully. External Data Sources Users can also access data from external databases for model training. For external systems follow the following steps: Click the Data Upload tab. Click External Data Sources.

Click External Data Sources. Select different sql databases and data warehouses to ingest data. viz.- SQLite, Microsoft SQL, PostgreSQL, MySQL, Oracle, Google BigQuery, Redshift, Snowflake, Actian. To access data from the database, you need to provide essential credentials like Database Name, Username, Host Port, etc. If a single table contains all the data, you can simply put the table name and fetch the table as Data Frame. Otherwise, the user can perform join and selection based on the condition to merge more than one table and finally fetch the data.  Note: As of now, this option is to upload data from some external databases and some data warehouses like- SQLite, PostgreSQL, etc. Users can insert a single table as a database. Also, some basic SQL operation like join, and selection based on where is also possible on multiple tables. Please contact ERS Research in case you wish to enable this option. SKLearn Data Sources To upload a standard sklearn data set, follow the steps given below: Click on the Data Upload tab. Click SKLearn Data Sources. Select a dataset. Click Submit. As a result, the dataset will be uploaded successfully. AION Labelled Data Here the trained unsupervised model can be used as a base model to train the supervised model. The two main significance of using a model within model is to simplify a large, complex model and allow for more advanced use of model iterators. To use the model chain following steps are followed: Submit new use case  Upload dataset Train unsupervised model i.e., clustering   Create a new use case  upload the dataset. Training a supervised model i.e., classification, regression, etc. From the drop-down list choose the model as the base model to train the new supervised model. Click next. EDA EDA stands for Exploratory Data Analysis where user can analyze and investigate data sets and summarize their main characteristics, employing data visualization methods. After uploading the data file, you can view the dataset of top 10 records.

To see the remaining details, click the EDA button. On clicking the EDA button, a window will pop up that enables configuring the features and the sample of data that needed to be selected for EDA. Select features and data subsampling size for exploratory data analysis. Finally, click Next. Data Overview: The summary statistics of the features can be viewed as shown in the figure below. Data Distribution: It shows the data distribution type for each feature of a dataset. For numerical data, select the numerical feature and click on Show Data Distribution to see the distribution of data is generated as shown below. For text data, select the text feature and click on the Show Data Distribution. Hence, word cloud is generated as shown below. More details can be seen by clicking the icon on the right of the Data Distribution tab. Timeseries Analysis: Check the dataset Stationarity and Seasonality for each of the feature of the ingested dataset. Feature Importance: The relevant features depending on the variance ratio can be viewed in the bar chart. The higher the variance ratio more relevant the feature is. For example: From the figure below, the sepal_width is the most significant feature for iris data classification. Correlation Analysis: The strength of relationships varying from 0 to 1 among various features can be viewed as shown in the figure below. A value closer to 1 shows that the features are highly correlated and a value close to 0 shows that the features are less correlated. Example: The correlation value of feature row record_dateTime and feature column id is 1 which shows that these two features are highly correlated and considering both the feature is irrelevant. Unsupervised Clustering: The unsupervised clustering of the data can be viewed as shown in the figure. Example: From the figure, we observed five clusters are being created and the average value of those clusters for feature retransmission is 5.52, 3.98, 5.23, 4.66 & 5.44.

Data Deep Dive: It offers an interface to inspect the correlation between data points for all the features of a dataset. Each item in the visualization represents the data point of an ingested dataset. Clicking on an individual item shows the key pairs that represent the features of that record whose values may be strings or numbers. Data deep dive related details will appear only when the User clicks the Show Data Deep Dive link to see it. Note: If the data ingestion is huge, Data Deep Dive may not be responsive, or data may fail to load. Pair Graph: The variation of two selected features can be compared as shown in the figure below. Fairness Metrics: The bias identification of data can be viewed by the fairness metrics. There are three metrics- Statistical Parity Difference, Disparate Impact and Theil Index to check for unwanted bias in the dataset. Both are measures of discrimination (i.e., deviation from fairness). Statistical Parity measures the difference that the majority and protected classes receive a favorable outcome. This measure must be equal to 0 to be fair. Disparate impact compares the proportion of individuals that receive a favorable outcome for two groups, a majority group, and a minority group. This measure must be equal to 1 to be fair. Select the feature (Categorical) and target feature and then click Show Metrics to analyze the potential bias between the selected feature and the target variable. Theil Index measures inequality among individuals in the group for a classification problem type. Ideally it should be zero for perfect fairness and biasness increase with increase in value greater than zero. Note: The acceptable threshold is highlighted in green. A class is considered biased if the metrics score is outside the threshold range. Basic Configuration This section provides the option to select more than one algorithm based on problem type and also provides the option to discard the unnecessary features from the dataset. Problem Description Model features

Problem Description Model features Filter and Grouping Problem Description Select the Problem Type. The following problem types are supported: Classification Regression Time Series Forecasting Time Series Anomaly Detection Recommender System Clustering Anomaly Detection Survival Analysis Topic Modeling State Transition Document Similarity Text Summarization [beta] Note: By default, classification is selected. For AWS EC2 (LLaMA 7B) and GCP GCE (LLaMA 13B) computer infrastructure, problem type Large Language Models-Tuning is available. Select the Scoring Criteria. Based on the problem Type, select the Algorithms. Select the Online Learning (Beta) or Distributed learning (Beta) check box for Incremental Learning. Only Regression and Classification problem types support Incremental learning. For more details, please refer to the sections later these topics. Features Configuration This section provides the option to select features that will be used for training, options to change the data type and transform each selected feature. Note that the target feature is not considered a training feature. In this section, the user can see the list of features with prefilled default values. Type of feature can be selected as date, index, categorical, numerical, or text. Six fill method is supported with the default value for numerical being Median and for Categorical is Mode. User can select different Encoding techniques for different features. Outlier and Normalization can be selected from the list of given options. Finally mention the Target Features (for supervised learning). For some of the algorithms like Document similarity and clustering, the target feature is not applicable. Time Series Settings Time Series Settings is for time series problem type, that allows to set the parameters like- Interpolation, Rolling Window, Aggregation, Granulatity, etc. Row Filtering and Grouping This section handles the row filtering. If filtering is not required, this section can be skipped.

Clicking on the Next button will save the configuration and moves it to Advanced Configuration Tab. AION Online Learning (Beta) Online learning is a method of machine learning in which input data is continuously used to extend the existing model's knowledge i.e., to further train the model even after the first fit. In AION, support for incremental learning is provided for regression and classification tasks. Every Online learning use case has two major components, Online data profiling - River and Sklearn Online model training – Sklearn Since both profiler and ml models must be continuously updated, the models are pickled and saved in the usecase folder for further use. Note that hyperparameter tuning can be done only in the first fit of data and for all the remaining incremental addition of data the selected best hyperparameters from the first fit are used. Instructions for creating a usecase: STEP 1: In the Basic, Config page click on the Online Learning checkbox and select the required options. STEP 2: Select the required data profiling options – Numerical fill – The method by which missing values in numerical features are filled Mean – Running Mean Min – Running Minimum Max – Running Maximum Drop – Drop the missing value records Zero – Fill with 0 Categorical fill – The method by which missing values in categorical features are filled Mode – Running Mode Drop – Drop the missing value records Zero – Fill with 0 Category encoding – The method by which categorical features are to be numerically encoded. Only Target encoding is supported. Target encoding params – The statistics to be used for encoding the target column into a categorical column are: Mean – Mean of the target group concerning the category is computed Bayesian Mean – A smoothed mean statistic to prevent data leakage while encoding Outlier detection – The method to use to remove outliers from training data are: One class SVM with a quantile filter Half space trees with quantile filter

Half space trees with quantile filter Other profiling options are the same as in regular AION profiler STEP 3: Train the model Now a usecase folder will be created in the target folder. All the profiler models and the best ML model are present in the Production folder Path: C:\Users\<username>\AppData\Local\HCLT\AION\target\<usecasename> \Production STEP 4: Incremental addition of new data to the existing model When new data is fed, the older model is archived in archives and the updated model is stored in the production folder. Path: C:\Users\<username>\AppData\Local\HCLT\AION\target\<usecasename> \archives The next fit or incremental fit logs are appended to the main model training logs. Users can check what profiling steps are applied and the model scores at each partial fit in the appended logs. After training the model with the first dataset, click the online batch learning icon from the Train tab for incremental learning. Pop-up window will appear to enter the incremental data (second or third dataset) location consisting of similar characteristics. Click on submit button. Accuracy results will pop up on the window with previous and current accuracy. STEP 5: Prediction of test data from the current production model Click on Predict icon from the predict tab. Predictions can be either from a single instance (Predict Instance) or from a file (Predict Batch). Click Predict or Submit for successful prediction. Limitations of Online Learning (Beta) For the classification task, the user is expected to not add new classes of the target variable in the next batch of data. Ensure that the next batch of data does not miss out any columns from previous data columns. Text features are not supported. Only Model training and prediction are supported. Distributed Learning (Beta)

Distributed Learning (Beta) Distributed machine learning refers to multi-node/multi-processor machine learning algorithms and systems that are designed to improve performance, increase accuracy, and scale to larger input data sizes. In AION, distributed learning is supported for classification, and regression with XGBoost and LightGBM. Characteristics of Distributed learning: Distributed data pre-processing and EDA: AION support loading, preprocessing, and analyzing a large dataset that exceed the memory limit of host machine. Provide advanced configuration for the user to easily configure the distributed parameters as well as model hyper-parameters. Distributed model training: Use the native XGBoost and LightGBM library for training and evaluating the model. Instructions for creating a use case STEP 1: In the Basic, Config page click on the Distributed Learning checkbox and select the algorithm. Distributed learning supports both classification and regression problems for algorithms- XGBoost and LightGBM. STEP 2: Select the model features and target features. STEP 3: Click the advance configuration tab to set the advance configuration or else click Next to train the model. STEP 4: Click the Train Model tab to train the model successfully. FastText Support For Document Similarity problem if FastText algorithm is selected than the user needs to be installed FastText pretrained model manually. Following are the steps to install the same: Download and install the Microsoft C++ Build Tools through link: https://visualstudio.microsoft.com/visual-cpp-build-tools Open the AION shell. Install fastText by using the command below- Command: python -m pip install fastText. or When it throws an error, Download whl package from link: https://www.lfd.uci.edu/~gohlke/pythonlibs/#fasttext Command: python -m pip install (Path of downloaded whl package) VectorDB Cosine For VectorDB cosine user needs to install the chromadb manually. Following are the steps to install the same:

Following are the steps to install the same: Download and install the Microsoft C++ Build Tools through link: https://visualstudio.microsoft.com/visual-cpp-build-tools Open the AION shell. Install chromadb by using the command below- Command: python -m pip install chromadb Advance Configuration To update the advanced configuration, follow the steps given below: Followed by data uploading and basic configuration click on the Advance Configuration tab. Configuration can be done for Data Filter, Text Transformations, and Feature Engineering, Machine Learning. Click on plus (+) icon for detail configuration. By default, the fields are pre-filled. Change the fields if required. Click Next and as a result, the advance configuration details are saved. Text embedding Text embedding measures the relatedness of text strings. Also, it has an ability to capture semantic meanings and relationships between words or phrases which enables machines to understand and process human language efficiently. Popular models to perform text embeddings are- TF_IDF, OpenAI, MiniLM-L6-v2 and Mpnet-base-v2. Model Training Model training tab gives the information about the Problem type, Selected algorithms, no of rows and Training Features, Training Data, Data balancing and Number of iterations. The data balancing is applicable only for classification problem. The percentage of test data can be changed accordingly by the user. Number of iterations reflects the size of hyper-parameter space selected by the user. Click Train Model tab to train the model. Training can be aborted if required by clicking Stop Training tab. Model training can also be performed by using API which is explained in detail in the section 5.1. Trained Model Details After successful training, tained model details can be viewed. Trained Model Info

Trained Model Info It gives an information about- Training status, Deployment location of a model, problem type, best model (in case multiple algorithms is selected), Score of a best model, feature used, and the reports can be downloaded by clicking on Download Reports tab. Evaluated Models Evaluated models gives an information about the model performance based on testing score, Confidence Score and Feature Engineering Method. For classification and regression performance of seclected model is compared with the basic model such as logistic regression and linear regression respectively. Eg: When user selects Extreme Gradient Boosting (XGBoost) model for classification, its performance will be compared with Logistic regression. Training Data evaluation It gives an information about training data as shown in the figure below. Test data evaluation It gives an information about test data as shown in the figure below. Performance Performance tab shows the graph of True Positive Rate vs False Positive Rate and Precision vs Recall, giving an idea about the model performance. Performance Leaderboard Performance leaderboard gives an idea about the best performing algorithm when multiple algorithms are selected based on its best score. Best performing algorithms appear at the first row followed by least performing algorithms. It supports only machine learning algorithms for classification and regression problem types. To view the leaderboard, click on the icon from the right corner of the model training window. Leaderboard will appear as shown in the figure below. Model Prediction Model prediction predicts the input data based on the pattern of training data. Predict Instance To perform the Single Instance prediction, follow the steps given below: Click on the Predict icon. If necessary, change feature values. Click Predict to obtain the prediction result. Predict Batch To perform the prediction from the file, follow the steps given below: Click on the Predict icon.

Click on the Predict icon. Click Predict Batch tab. Upload the dataset. Click Submit to view the prediction result for the file selected. Prediction can also be performed by using API which is explained in detail in the section 5.2. Generate Script This will generate the sample python code to call rest API for prediction. Note: Prediction Service in AION is only for testing purposes and not for performance. For performance, the prediction service must be published. Trusted AI Trusted AI provides an integrated approach to evaluate, quantify, and monitor the impact and trustworthiness of a models. Fairness Metrics Fairness metrics are a set of measures that enables the user to detect the presence of bias in model. Theil Index measures inequality among individuals in the group for a classification problem type. Value of Theil Index is always greater than zero. Ideally it should be zero for perfect fairness and biasness increase with increase in value higher than zero. Equal Opportunity Difference measure the difference in true positive rate between privileged and unprivileged group of protected features for a classification problem type. Ideally it should be zero to be fair. A value less than zero illustrates a higher benefit for the privileged group and positive value is a sign of higher benefit for the unprivileged group. Statistical Parity Difference measure the difference in the rate of favorable outcomes between privileged and unprivileged groups. This measure must be equal to 0 to be fair. Disparate impact ratio measures the ratio of rate of favorable outcomes for the underprivileged group to that of privileged group. This measure must be equal to 0 to be fair. Steps: Click on trusted AI icon under predict tab from the home page. Select protected feature from drop down column ‘Check Bias For’. Target feature gets selected automatically which is selected while training the model. Select the graph type. Click Show Metric tab to view the graph.

Click Show Metric tab to view the graph. Note: The acceptable threshold is highlighted in green. A class is considered biased if metrics-score is outside the threshold range. Performance Model performance is an assessment of the model's ability to perform a task accurately not only with training data but also in real-time with runtime data when the model is deployed. Matthews correlation coefficient (MCC) is generally regarded as a balanced measure which can be used even if the classes are of very different sizes. The MCC is in essence a correlation coefficient value between -1 and +1. A coefficient of +1 represents a perfect prediction, 0 an average random prediction and -1 an inverse prediction. Transparency Transparency helps users to understand the input data, algorithm and expected output of an ML model. Brier Score measures the mean of squared difference between the predicted probability between the model and the ground truth values of a target feature. Explainability Explainability tab gives an explanation about Prediction and Model. Xplain Prediction Xplain Prediction explain the details about prediction concerning its Precision, confidence, reason, along with force plot and decision plot. Steps: Click on Trusted AI icon. Click Xplain Prediction. Click Xplain. Detailed Result will be displayed as shown below. Xplain Model Xplain Model gives the insights of input data and explain the details of its important features. Steps: Click on Trusted AI icon. Click Xplain Model tab. Click Get Model Details. The details of the model explanation will be displayed. Scroll down for more details. User can DOWNLOAD REPORTS if required. Robustness Robustness of AI/ML models is analyzed by Sensitivity Analysis examining the extent to which the results are affected by changes in input data. We have two different options to perform sensitivity analysis to understand the influencing factors in more details. First Order Sensitivity Analysis