Evaluating uncertainty-based active learning for accelerating the generalization of molecular property prediction

Table 4 Redox potential uncertainty estimation evaluation

UQ Type	Method	Model	ENCE	\(\rho _{\text {error}}\)	\(\rho _{\text {ood}}\)	\(\rho _{\Delta \text {error}}\)
Baseline	GBM	GBM	0.524	\(\mathbf {0.283 \pm 0.001}\)	0.296	0.679
Ensemble	MCDO	MDM	1.742	0.246	0.329	0.814
	MCDO	GNN	1.428	0.112	0.346	−0.461
	Ensemble	MDM	1.730	0.265	0.407	0.225
	Ensemble	GNN	2.417	0.028	0.405	0.229
Target value	Evidential	MDM	7.441	0.386	0.103	−0.125
	Evidential	GNN	2.718	0.013	0.389	0.054
	MVE	MDM	\(\mathbf {0.273 \pm 0.066}\)	0.384	0.278	0.300
	MVE	GNN	\(\mathbf {0.074 \pm 0.006}\)	0.034	0.230	−0.118
Union	GBM	MDM	0.455	0.191	0.165	−0.296
Union	GBM	GNN	0.466	0.195	−0.118	−0.071
Distance	Density (EB)	MDM	–	0.199	0.578	0.743
		GNN	–	0.162	0.578	0.743
		GBM	–	0.119	0.578	0.693

The uncertainty approach with best mean performance across models for each metric is shown in bold. For the density method, EB refers to embedding-based similarity

ISSN: 1758-2946