Table 3. Open in new tab Ablation...

QB	BR	AvUC	Tox21	SIDER	MUV
✗	✗	✗	0.815	0.821	0.733	0.720	0.630	0.552
			11%	8%	41%	43%	0%	0%
✓	✗	✗	0.816	0.817	0.698	0.698	0.628	0.638
			62%	54%	73%	68%	23%	30%
✓	✓	✗	0.832	0.833	0.776	0.783	0.629	0.631
			64%	60%	76%	70%	29%	40%
✓	✓	✓	0.833	0.834	0.792	0.794	0.629	0.630
			53%	50%	74%	69%	28%	35%

QB	BR	AvUC	Tox21	SIDER	MUV
✗	✗	✗	0.815	0.821	0.733	0.720	0.630	0.552
			11%	8%	41%	43%	0%	0%
✓	✗	✗	0.816	0.817	0.698	0.698	0.628	0.638
			62%	54%	73%	68%	23%	30%
✓	✓	✗	0.832	0.833	0.776	0.783	0.629	0.631
			64%	60%	76%	70%	29%	40%
✓	✓	✓	0.833	0.834	0.792	0.794	0.629	0.630
			53%	50%	74%	69%	28%	35%

Table 3.

Ablation studies of EM3P2 reporting ROC-AUC (floating point) and precision (percentage).^a^,^b

QB	BR	AvUC	Tox21		SIDER		MUV
QB	BR	AvUC	1-shot	10-shot	1-shot	10-shot	1-shot	10-shot
✗	✗	✗	0.815	0.821	0.733	0.720	0.630	0.552
			11%	8%	41%	43%	0%	0%
✓	✗	✗	0.816	0.817	0.698	0.698	0.628	0.638
			62%	54%	73%	68%	23%	30%
✓	✓	✗	0.832	0.833	0.776	0.783	0.629	0.631
			64%	60%	76%	70%	29%	40%
✓	✓	✓	0.833	0.834	0.792	0.794	0.629	0.630
			53%	50%	74%	69%	28%	35%

QB	BR	AvUC	Tox21		SIDER		MUV
QB	BR	AvUC	1-shot	10-shot	1-shot	10-shot	1-shot	10-shot
✗	✗	✗	0.815	0.821	0.733	0.720	0.630	0.552
			11%	8%	41%	43%	0%	0%
✓	✗	✗	0.816	0.817	0.698	0.698	0.628	0.638
			62%	54%	73%	68%	23%	30%
✓	✓	✗	0.832	0.833	0.776	0.783	0.629	0.631
			64%	60%	76%	70%	29%	40%
✓	✓	✓	0.833	0.834	0.792	0.794	0.629	0.630
			53%	50%	74%	69%	28%	35%

QB, query balancing in meta training; BR, Belief Regularizer; AvUC, Accuracy Versus Uncertainty Loss Regularizer.

The precision values are calculated for the minority labels.

The best performances are in bold face numbers.