AUC maximizing support vector machines with feature selection

Abstract

In this paper, we proposed a new algorithm, the Sparse AUC maximizing support vector machine, to get more sparse features and higher AUC than standard SVM. By applying p-norm where 0 < p < 1 to the weight w of the separating hyperplane (w · x) + b = 0, the new algorithm can delete less important features corresponding to smaller |w|. Besides, by applying the AUC maximizing objective function, the algorithm can get higher AUC which make the decision function have higher prediction ability. Experiments demonstrate the new algorithm's effectiveness. Some contributions as follows: (1) the algorithm optimizes AUC instead of accuracy; (2) incorporating feature selection into the classification process; (3) conduct experiments to demonstrate the performance.