Improving SVM classification on imbalanced datasets by introducing a new bias.

Abstract

Support Vector Machine (SVM) learning from imbalanced datasets, as well as most learning machines, can show poor performance on the minority class because SVMs were designed to induce a model based on the overall error. To improve their performance in these kind of problems, a low-cost post-processing strategy is proposed based on calculating a new bias to adjust the function learned by the SVM.

The proposed bias will consider the proportional size between classes in order to improve performance on the minority class. This solution avoids not only introducing and tuning new parameters, but also modifying the standard optimization problem for SVM training.

Experimental results on 34 datasets, with different degrees of imbalance, show that the proposed method actually improves the classification on imbalanced datasets, by using standardized error measures based on sensitivity and g-means. Furthermore, its performance is comparable to well-known cost-sensitive and Synthetic Minority Over-sampling Technique (SMOTE) schemes, without adding complexity or computational costs.