Discriminant Feature Selection via Minimizing the Minority-emphasis Regularization to Within-class Matrix for Imbalanced Data

Abstract

In real-world applications, imbalanced data is frequently found, such as medical diagnosis and intrusion detection. Due to the hypothesis that samples are evenly distributed among classes, the classification performance obtained by traditional classifiers is easily influenced by class-imbalance issue. Feature selection is an important way to retain a more discriminative feature subset with less redundancy. The performance on minority class needs to be focused in feature selection for imbalanced data. In this study, to address class-imbalance issue, a new form of regularization to LDA (IR-LDA) is embedded into a feature selection framework, in which $\ell_{2}-$ norm regularization is adopted. A feature selection algorithm (IR-DFS) is presented. An efficient optimization algorithm is applied to achieve the optimal solution. The effectiveness of IR-DFS is demonstrated by experiments on public imbalanced data sets as well.