Table 1.

Evaluation of GNN explainers on SG-Base graph dataset based on node explanation masks ${\mathbf{M}}_N^p$. Arrows (↑/↓) indicate the direction of better performance.

Method	GEA (↑)	GEF (↓)	GES (↓)	GECF (↓)	GEGF (↓)
Random	0.148 ± 0.002	0.579 ± 0.007	0.920 ± 0.002	0.763 ± 0.003	0.023 ± 0.002
Grad	0.193 ± 0.002	0.392 ± 0.006	0.806 ± 0.004	0.159 ± 0.004	0.039 ± 0.003
GradCAM	0.222 ± 0.002	0.452 ± 0.006	0.263 ± 0.004	0.010 ± 0.001	0.020 ± 0.002
GuidedBP	0.194 ± 0.001	0.557 ± 0.007	0.432 ± 0.004	0.067 ± 0.002	0.021 ± 0.002
IG	0.142 ± 0.002	0.545 ± 0.007	0.727 ± 0.005	0.110 ± 0.003	0.021 ± 0.002
GNNExplainer	0.102 ± 0.003	0.534 ± 0.007	0.431 ± 0.008	0.233 ± 0.006	0.027 ± 0.002
PGMExplainer	0.133 ± 0.002	0.541 ± 0.007	0.984 ± 0.001	0.791 ± 0.003	0.096 ± 0.004
PGExplainer	0.194 ± 0.002	0.557 ± 0.007	0.217 ± 0.004	0.009 ± 0.000	0.029 ± 0.002
SubgraphX	0.324 ± 0.004	0.254 ± 0.006	0.745 ± 0.005	0.241 ± 0.006	0.035 ± 0.003

SubgraphX far outperforms other methods in accuracy and faithfulness while PGExplainer is best for stability and counterfactual fairness. In general, gradient methods produce the most fair explanations across both counterfactual and group fairness metrics. See Tables 3–4 for results on edge and feature explanation masks.