Towards Decision-Friendly AUC: Learning Multi-Classifier with AUCµ

Area Under the ROC Curve (AUC) is a widely used ranking metric in imbalanced learning due to its insensitivity label distributions. As well-known multiclass extension of AUC, Multiclass AUC (MAUC, a.k.a. M-metric) measures average multiple binary classifiers. In this paper, we argue that simply optimizing MAUC far from enough for multi-classification. More precisely, only focuses on scoring functions via optimization, while leaving decision process unconsidered. Therefore, being able make good decisions might suffer low performance terms MAUC. To overcome issue, turn explore AUCµ, another variant which further takes into consideration. Motivated by fact, propose surrogate risk optimization framework improve model perspective AUCµ. Practically, two-stage training multi-classification, where at first stage function learned maximizing and second seek F1-metric our proposed soft F1. Theoretically, provide sufficient conditions losses could lead Bayes optimal function. Afterward, show enjoys generalization bound order O(1/√N). Experimental results four benchmark datasets demonstrate effectiveness method both AUCµ F1-metric.