Table 5.

Explanation consistency, C, of checkpoint averaging, submodel averaging and random submodel picking on baseline models and both normal and explanation ensembles. CA-SA is checkpoint averaging followed by ensemble submodel averaging, CU the Codon Usage dataset, class. is classification and regr. regression.

		BCW	Diabetes (Class.)	Diabetes (Regr.)	CU (DNA)	CU (Kingdom)	MIMIC-IV
Baseline models	Checkpoint averaging	0.2117	0.75322	0.53356	0.06585	0.06378	0.1527
Normal ensemble models	Checkpoint averaging	0.2497	0.2790	0.5604	0.0007	0.2264	n/a
	Random submodel	0.1392	0.3062	0.5075	0.0440	0.01722	n/a
	Submodel averaging	0.1952	0.4597	0.4713	0.3882	0.1738	n/a
	CA-SA	0.2906	0.551	0.5193	0.5654	0.1638	n/a
Explanation ensemble models	Checkpoint averaging	0.2485	0.0175	0.5322	0.2510	0.2695	0.2954
	Random submodel	0.1365	0.2641	0.0625	0.2939	0.0193	0.01333
	Submodel averaging	0.2983	0.0355	0.8389	0.0953	0.0330	0.2080
	CA-SA	0.3964	0.89222	0.8529	0.6462	0.3481	0.1784