Label correlation in multi-label classification using local attribute reductions with fuzzy rough sets

Abstract

To improve the efficiency and effectiveness of multi-label learning tasks, label correlation has been explored and widely applied in multi-label classification. However, the existing common way to obtain label correlation is to count the co-occurrence frequency of labels in output space. From our perspective, the feature variables of input space also contain the basic information, which can affect each label variable in output space. Obviously, the crucial features for label can efficiently reflect the internal relationship between input space and the label variable. The high consistency of the crucial features for different labels indicates that these labels are closely related. Inspired by the above intuitive motivations, the crucial features explicitly for every local label class are defined and calculated based on local attribute reductions with fuzzy rough sets. Local label correlations are naturally constructed, and then global label correlation is obtained by aggregating related local label correlations. According to different judgment parameters α and global label correlation, the labels can be transformed into several independent label-related subsets. To search for the indispensable and representative features of each label-related subset, the local-reduction-based feature selection method (LRFS-α) is designed. Comprehensive experimental results on fifteen multi-label datasets characterize the performance of our methods against other seven multi-label learning methods.