Enhanced Maximum AUC Linear Classifier
Xiannian Fan, Ke Tang · 2010 Seventh International Conference on Fuzzy Systems and Knowledge Discovery · 2010
In the field of imbalance learning and cost sensitive learning, minimization of the classification error rate is not an appropriate approach due to class skew and cost distributions. Thus the area under the ROC Curve (AUC) has been widely utilized to assess the performance of the classifiers in such cases. The Maximum AUC Linear Classifier (MALC), aiming at maximizing AUC directly, is a nonparametric linear classifier. MALC is based on the analysis of Wilcoxon-Mann-Whitney statistic of each single feature and on greedy pairwise combinations of the features. This paper finds that the MALC searches the solution in a much constrained resolution space. Furthermore, the heuristic method for guiding the structure of the classifier is worthy of notice. In this paper the Enhanced MALC (EMALC) is proposed. In the EMALC, two modifications are presented. Modification 1 aims at extensive searching in the solution space. Modification 2 modifies the way that MALC guides to induce the structure of the classifier. Experimental studies are carried out on a broad range of real world dataset. And the proposed methods have shown significant effect.