Table 3.

Area under the curve (AUC) of interpretability methods for each model and each classification performance metric evaluated using ROAR. AUC is measured for two prediction metrics (AURPC and AUROC) respectively. Lower AUC indicates more rapid prediction performance drop and better feature importance interpretation.

Interpreters	AutoInt		LSTM		TCN		Transformer		IMVLSTM
	AUPRC	AUROC	AUPRC	AUROC	AUPRC	AUROC	AUPRC	AUROC	AUPRC	AUROC
Random	0.401	0.842	0.615	0.909	0.605	0.901	0.662	0.918	0.669	0.915
Glassbox	×	×	×	×	×	×	×	×	0.533	0.892
Saliency	×	×	0.558	0.898	0.587	0.893	0.616	0.909	0.566	0.884
IntegratedGradients	×	×	0.586	0.899	0.593	0.899	0.588	0.903	0.465	0.863
DeepLift	×	×	0.575	0.900	0.598	0.898	0.594	0.905	0.542	0.883
GradientShap	×	×	0.561	0.893	0.592	0.899	0.600	0.904	0.470	0.858
DeepLiftShap	×	×	0.569	0.897	0.607	0.901	0.619	0.909	0.554	0.887
SaliencyNoiseTunnel	×	×	0.551	0.892	0.581	0.896	0.578	0.899	0.475	0.851
ShapleySampling	0.456	0.866	0.628	0.910	0.613	0.898	0.655	0.916	0.668	0.917
FeaturePermutation	0.454	0.866	0.624	0.910	0.616	0.903	0.655	0.917	0.677	0.918
FeatureAblation	0.279	0.733	0.438	0.811	0.479	0.824	0.425	0.792	0.408	0.830
Occlusion	0.456	0.866	0.617	0.909	0.609	0.898	0.653	0.917	0.684	0.920
ArchDetect	0.251	0.696	0.369	0.774	0.446	0.818	0.379	0.784	0.382	0.805