A Feature Selection Method for Multivariate Performance Measures

Feature selection with specific multivariate performance measures is the key to the success of many applications such as information retrieval. In this paper, we propose a feature selection method for multivariate performance measures. The proposed method forms an optimization problem with exponential size of both feature groups and label configurations for a given dataset. To address this problem, a two-layer cutting plane algorithm is proposed. The outer layer performs group feature generation; while the inner layer learns the label configuration for multivariate performance measures. Comprehensive experiments on large-scale and high-dimensional real world datasets show that the proposed method can significantly outperform $l_1$-SVM and SVM-RFE when choosing a small subset of features, and achieve significantly improved performances over SVM$^{perf}$ in terms of $F_1$-score. It also learns a sparse yet effective decision rule for multivariate performance measures.