Multi-Label Learning Through Label-Specific Features with Entropy Definition

Multi-label learning deals with the problem where each instance is associated with multiple labels. To exploit discriminative features for different labels, some researchers seek to construct label-specific features for model induction. In this scheme, clustering analysis is generally conducted for label-specific features generation, and clustering ensemble can be employed to enhance the exploration on label correlation. However, previous work fail to effectively consider the distribution information of multi-label data, and thus the performance of clustering analysis is limited. To address this issue, we propose a novel method named LIFTED, which systematically employs information theory to guide clustering and clustering ensemble. First, for distinct label, an objective function is established to determine the degree of clustering through minimizing the label entropy. Second, an entropy-based label similarity is designed to guide clustering ensemble, which enhances the model stability and its ability on label correlation exploiting. Experiments on 12 benchmark datasets validate the competitive performance of the proposed method against the state-of-art algorithms.