IC-1044Counterfactual GNN explainers produce statistically infeasible recourses that violate topological constraints in molecular datasets

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 tests whether counterfactual explanations generated for molecular graphs correspond to valid molecules by checking connectivity. In the original datasets, molecules are expected to be single connected components. The paper measures the p-value of the deviation between the expected count of connected graphs and the observed count in the counterfactuals. For RCExplainer and CF2 on mutagenicity and AIDS, the deviations are highly significant (p < 0.00001), meaning the counterfactuals frequently produce disconnected molecules. Visualization in Appendix F further shows that CLEAR and CF2 produce explanations violating valency rules.

Evidence
observational
Key metric
Mutagenicity: RCExplainer expected 233.05 connected, observed 70, p < 0.00001; CF2 expected 206.65, observed 0, p < 0.00001. AIDS: RCExplainer expected 17.6, observed 8, p < 0.00001; CF2 expected 1.76, observed 0, p = 0.0001 (Table L)
Caveat
Feasibility is only tested on three molecular datasets (mutagenicity, MUTAG, AIDS). MUTAG shows no significant deviation (p = 0.55 for RCExplainer, p = 0.13 for CF2), so the issue is dataset-dependent.
Model
RCExplainer, CF2, CLEAR
Concepts
Failure mode
Datasets
MUTAG [eval], AIDS [eval]
Related findings
IC-1045
Extraction
automatic-extraction