Multi-label Classification via Label-Topic Pairs

The task of learning from multi-label example is rather challenging because of the tremendous number of possible label sets. It has been well recognized that exploiting label relationships in a proper way can facilitate the learning process and boost the learning performance. In this paper, we propose a novel framework called Label-Topic Pairs Multi-Label (LTPML) for multi-label classification. LTPML regards the label set associated with each instance as a document and each class label in the label set as a word and then obtains the topics from the label space by topic models. With the information about label correlations contained by topics, multi-label classification problem is decomposed into a series of single-label classification problems. Based on label-topic pairs which are constructed from relationships among the current label and topics, several multi-class classifiers are built for each class label. Two algorithms named LTPML-alpha and LTPML-beta are derived according to different way of selecting the topics. Experiments on benchmark data sets clearly validate the effectiveness of the proposed approaches.