Multi-sample $$\zeta $$ -mixup: richer, more realistic synthetic samples from a p-series interpolant

Journal of Big Data

Table 7 Classification performance (AUC and ACC) evaluated on MEDMNIST

Dataset	PathMNIST		DermaMNIST		OCTMNIST		PneumoniaMNIST		BloodMNIST
#images (#classes)	107,180 (9)		10,015 (7)		109,309 (4)		5,856 (2)		17,092 (8)
Method	AUC	ACC	AUC	ACC	AUC	ACC	AUC	ACC	AUC	ACC
ERM	0.962	84.4%	0.899	72.1%	0.951	70.8%	0.947	80.3%	0.995	92.9%
mixup	0.959	77.5%	0.897	72.2%	0.945	70.5%	0.945	75.4%	0.994	94.4%
$\zeta $-mixup ($\gamma =2.8$)	0.969	87.6%	0.911	73.3%	0.918	72.8%	0.951	80.9%	0.997	95.2%

Dataset	TissueMNIST		BreastMNIST		OrganMNIST_A		OrganMNIST_C		OrganMNIST_S
#images (#classes)	236,386 (8)		780 (2)		58,850 (11)		23,660 (11)		25,221 (11)
Method	AUC	ACC	AUC	ACC	AUC	ACC	AUC	ACC	AUC	ACC
ERM	0.911	62.7%	0.897	85.9%	0.995	92.1%	0.990	88.9%	0.967	76.2%
mixup	0.911	63.2%	0.914	76.2%	0.995	93.1%	0.990	89.9%	0.966	72.7%
$\zeta $-mixup ($\gamma =2.8$)	0.918	63.9%	0.928	87.2%	0.996	92.7%	0.991	91.0%	0.969	77.1%

The evaluation metrics are the area under the ROC curve (‘AUC’) and the classification accuracy (‘ACC’). Higher values are better for both the metrics. The highest values of each metric have been formatted with bold