Graph Inference Data Valuation Framework

This repository contains code of Shapley-Guided Utility Learning for Effective Graph Inference Data Valuation.

Scripts Overview

1. preprocess.py

   • Preprocesses the graph dataset for GNN training and evaluation.
   • Input: Raw graph data
   • Output: Processed graph data ready for GNN consumption

2. train_gnn.py

   • Trains the GNN model on the preprocessed data.
   • Input: Processed graph data
   • Output: Trained GNN model

3. valid_perm_sample.py

   • Generates validation permutation samples for model evaluation.
   • Input: Trained GNN model, validation data
   • Output: Validation permutation samples

4. test_perm_sample.py

   • Generates test permutation samples for final model evaluation.
   • Input: Trained GNN model, test data
   • Output: Test permutation samples

5. atc_confidence_estimation.py

   • Estimates confidence using the Adaptive Test Confidence (ATC) method.
   • Input: Validation permutation samples
   • Output: ATC confidence estimates

6. atc_ne_confidence_estimation.py

   • Estimates confidence using the ATC with Negative Entropy (ATC-NE) method.
   • Input: Validation permutation samples
   • Output: ATC-NE confidence estimates

7. training_statistics.py

   • Computes and saves training set statistics for use in performance prediction.
   • Input: Trained GNN model, training data
   • Output: Training statistics

8. doc_performance_prediction.py

   • Predicts model performance using the Difference of Confidence (DOC) method.
   • Input: Training statistics, validation permutation samples
   • Output: DOC performance predictions

9. shapley_regression_pred_lasso.py

   • Performs Shapley regression prediction using LASSO regularization (our proposed method).
   • Input: Shapley values, performance metrics
   • Output: Regression model for performance prediction

10. shapley_estimation_drop_node.py

• Estimates Shapley values for nodes by dropping them from the graph.