A boosting method for direct AUC optimization

Zhongliang Li, Shaodan Zhai, Tian Xia, Shaojun Wang

Abstract: We present a boosting method for classification problems with optimal AUC value as a performance measure. The proposed technique first minimizes the empirical pairwise classification error. Once the pairwise classification error is reduced to a coordinatewise local minimum, then it switches to maximize the average pairwise margin of a small set of bottom sample pairs. Experimental results on real-world data sets show that the proposed non-convex optimization method achieves competitive or better results than the convex relaxation methods, and it is very robust in the noisy datasets.