Research map / Graph learning / GNN architectures
RegExplainer: Generating Explanations for Graph Neural Networks in Regression Tasks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RegExplainer, a post-hoc instance-level explainer for GNN graph regression that replaces the classification cross-entropy in the GIB objective with an InfoNCE-based lower bound and uses a graph mix-up with self-supervised contrastive learning to handle distribution shift.
Problem
Post-hoc instance-level explanation of GNN graph regression models is unexplored; existing GIB-based explainers rely on cross-entropy over classes and suffer distribution shift between explanation subgraphs and training graphs, which matters more for continuous decision boundaries.
Approach
Lower-bound the mutual information between explanation-subgraph predictions and labels with an InfoNCE loss over graph triplets, and build within-distribution mix-up graphs (explanation subgraph plus label-irrelevant remainder of a neighbor) so the GNN's embeddings and predictions are not evaluated on out-of-distribution subgraphs; the method is explainer-agnostic and plugs into GNNExplainer and P
Key results
- RegExplainer on PGExplainer reaches explanation AUC 0.758 on BA-Motif-Volume, 0.989 on BA-Motif-Counting, 0.739 on Triangles and 0.553 on Crippen, the best in all four datasets.
- Relative to the second-best baseline, RegExplainer (PGExplainer backbone) improves explanation AUC by 0.175 (34.3%) on average and up to 0.246 (48.0%).
- Prediction shift study shows RMSE of f(G*) vs Y is much larger than f(G) vs Y on BA-Motif-Volume (1432.07 vs 131.42), supporting the distribution-shift claim.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply RegExplainer framework to PGExplainer | PGExplainer · BA-Motif-Volume | AUC-ROC: 0.470 -> 0.758 | Helped |
| apply RegExplainer framework to PGExplainer | PGExplainer · BA-Motif-Counting | AUC-ROC: 0.798 -> 0.989 | Helped |
| apply RegExplainer framework to PGExplainer | PGExplainer · Triangles | AUC-ROC: 0.511 -> 0.739 | Helped |
| apply RegExplainer framework to PGExplainer | PGExplainer · Crippen | AUC-ROC: 0.448 -> 0.553 | Helped |
| remove mix-up processing (ablation) | RegExplainer with PGExplainer backbone · all four datasets | AUC-ROC | Hurt |
| remove InfoNCE loss while retaining mix-up and MSE loss (ablation) | RegExplainer with PGExplainer backbone · all four datasets | AUC-ROC | Hurt |
| remove MSE loss (ablation) | RegExplainer with PGExplainer backbone · all four datasets | AUC-ROC | Hurt |
| mix the explanation subgraph with a positive-neighbor graph's label-irrelevant subgraph | RegExplainer · BA-Motif-Volume | RMSE between original and explanation/mix-up predictions: 1427.07 -> 393.26 | Helped |
| mix the explanation subgraph with a positive-neighbor graph's label-irrelevant subgraph | RegExplainer · BA-Motif-Counting | RMSE between original and explanation/mix-up predictions: 7.22 -> 2.73 | Helped |
| mix the explanation subgraph with a positive-neighbor graph's label-irrelevant subgraph | RegExplainer · Triangles | RMSE between original and explanation/mix-up predictions: 12.40 -> 8.22 | Helped |
| mix the explanation subgraph with a positive-neighbor graph's label-irrelevant subgraph | RegExplainer · Crippen | RMSE between original and explanation/mix-up predictions: 1.17 -> 0.68 | Helped |
Surprising findings
- Plain MixupExplainer performs near chance on BA-Motif-Volume (AUC 0.471) and Crippen (0.499), below several simpler baselines.
- Explanation subgraph predictions shift substantially from the original predictions even when the original GNN predicts well, for example RMSE 1432 for f(G*) vs 131 for f(G) on BA-Motif-Volume.
Limitations
- Not applicable to spatio-temporal graphs with dynamic topology and node features, as acknowledged by the authors
- Evaluation relies on synthetic ground-truth explanations plus one small real-world dataset (Crippen), so real-world explanation quality is hard to verify
- Many reported results have large standard deviations (e.g., PGExplainer+RegExplainer 0.758 ± 0.177 on BA-Motif-Volume)
Experimental setup
Datasets: BA-Motif-Volume BA-Motif-Counting Triangles Crippen
Benchmarks: BA-Motif-Volume BA-Motif-Counting Triangles Crippen
Metrics: AUC-ROC on edges RMSE cosine distance Euclidean distance
Baselines: GRAD ATT GNNExplainer PGExplainer MixupExplainer
Builds on: GIB PGExplainer GNNExplainer MixupExplainer InfoNCE
Closest papers by meaning
- Factorized Explainer for Graph Neural Networks (AAAI 2024 · similarity 0.89)
- Towards Robust Fidelity for Evaluating Explainability of Graph Neural Networks (ICLR 2024 · similarity 0.89)
- Self-Interpretable Graph Learning with Sufficient and Necessary Explanations (AAAI 2024 · similarity 0.89)
- GOAt: Explaining Graph Neural Networks via Graph Output Attribution (ICLR 2024 · similarity 0.89)
- Stratified GNN Explanations through Sufficient Expansion (AAAI 2024 · similarity 0.88)
- Generating In-Distribution Proxy Graphs for Explaining Graph Neural Networks (ICML 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.