Table 2.

Evaluation of GNN explainers on SG-Base graph dataset based on node explanation masks ${\mathbf{M}}_N^p$.

Method	GEA (↑)	GEF (↓)	GES (↓)	GECF (↓)	GEGF (↓)
Random	0.075 ± 0.002	0.638 ± 0.007	1.55 ± 0.004	1.01 ± 0.010	0.027 ± 0.002
Grad	0.194 ± 0.002	0.498 ± 0.007	0.745 ± 0.005	0.157 ± 0.004	0.068 ± 0.003
GradCAM	0.188 ± 0.001	0.620 ± 0.006	0.295 ± 0.005	0.029 ± 0.003	0.027 ± 0.002
GuidedBP	0.190 ± 0.001	0.653 ± 0.007	0.430 ± 0.004	0.074 ± 0.003	0.020 ± 0.002
IG	0.140 ± 0.002	0.672 ± 0.007	0.639 ± 0.004	0.114 ± 0.004	0.011 ± 0.001
GNNExplainer	0.103 ± 0.003	0.632 ± 0.007	0.431 ± 0.008	0.249 ± 0.007	0.028 ± 0.002
PGMExplainer	0.133 ± 0.002	0.622 ± 0.007	0.974 ± 0.001	0.798 ± 0.003	0.083 ± 0.003
PGExplainer	0.165 ± 0.002	0.635 ± 0.007	0.224 ± 0.004	0.005 ± 0.000	0.030 ± 0.002
SubgraphX	0.383 ± 0.004	0.344 ± 0.006	0.585 ± 0.004	0.225 ± 0.004	0.114 ± 0.004

Base GNN is a GCN³³ as opposed to Table 1 which is based on explaining a GIN model³². Overall, explainer performance is very similar to that of the GIN with SubgraphX performing the best on faithfulness and accuracy metrics while gradient-based methods and PGExplainer typically perform best for fairness and stability.