A Label Correlation Based Weighting Feature Selection Approach for Multi-label Data

Abstract

Exploiting label correlation is important for multi-label learning, where each instance is associated with a set of labels. However, most of existing multi-label feature selection methods ignore the label correlation. Therefore, we propose a Label Correlation Based Weighting Feature Selection Approach for Multi-Label Data, called MLLCWFS. It is a framework developed from traditional filtering feature selection methods for single-label data. To exploit the label correlation, we compute the importance of each label in mutual information, and adopt three weighting strategies to evaluate the correlation between features and labels. Extensive experiments conducted on four benchmark data sets using two base classifiers demonstrate that our approach is superior to the state-of-the-art feature selection algorithms for multi-label data.