Table 5.

Area under the curve (AUC) of interpretability methods for each model and the model trustworthiness metric evaluated using ROAR. Lower AUC indicates more rapid trustworthiness drop.

Interpreters	AutoInt	LSTM	TCN	Transformer	IMVLSTM
Random	0.893	0.919	0.908	0.924	0.924
Glassbox	×	×	×	×	0.910
Saliency	×	0.911	0.914	0.922	0.914
IntegratedGradients	×	0.912	0.907	0.918	0.902
DeepLift	×	0.910	0.912	0.918	0.909
GradientShap	×	0.910	0.908	0.916	0.901
DeepLiftShap	×	0.911	0.914	0.921	0.911
SaliencyNoiseTunnel	×	0.915	0.916	0.913	0.900
ShapleySampling	0.902	0.924	0.917	0.920	0.925
FeaturePermutation	0.898	0.920	0.913	0.921	0.925
FeatureAblation	0.881	0.894	0.900	0.901	0.894
Occlusion	0.901	0.919	0.912	0.919	0.926
ArchDetect	0.880	0.892	0.894	0.894	0.891

Lower is better. Best values are in [bold].