IC-1045Factual GNN explanations do not capture the full data signal: retraining on explanations fails to reproduce predictions while retraining on residuals preserves them

Mert Kosan, Samidha Verma, Burouj Armgaan, Khushbu Pahwa, Ambuj Singh, Sourav Medya, Sayan Ranu

SourceGNNX-BENCH: Unravelling the Utility of Perturbation-based GNN Explainers through In-depth Benchmarking

The paper measures two reproducibility metrics. Reproducibility+ retrains a GNN on only the explanation subgraphs and compares test accuracy to the original model; values are consistently below 1.0, meaning the explanations alone are insufficient. Reproducibility- retrains a GNN on the residual graphs (original minus explanation) and finds accuracy remains close to the original, meaning the explanation edges are not the true cause of the prediction. The paper concludes that explainers capture only specific signals learned by the GNN and do not encompass all underlying data signals.

Evidence
correlational
Key metric
Reproducibility- stays near 1.0 across all methods and datasets (e.g. mutagenicity 0.99-1.06, AIDS 0.99-1.0, NCI1 0.89-0.93 in Table U); Reproducibility+ below 1.0 for most methods (e.g. mutagenicity 0.97-1.11, MUTAG 0.11-1.08 in Table T)
Caveat
SubgraphX results are unreliable due to very high variance from using only a small number of test graphs (stated in Table T and U captions). The base GNNs are small (3 layers, 20 hidden dims) and trained by the authors.
Model
PGExplainer, TAGExplainer, GEM, CF2, RCExplainer
Concepts
Explanation faithfulness
Datasets
MUTAG [eval], IMDB-B [eval], AIDS [eval], NCI1 [eval]
Related findings
IC-1044
Extraction
automatic-extraction