Multi-label learning based on label-specific features and local pairwise label correlation

Multi-label learning has drawn great attention in recent years. One of its tasks aims to build classification models for the problem where each instance associates with a set of labels. In order to exploit discriminative features for classification, some methods are proposed to construct label-specific features. However, these methods neglect the correlation among labels. In this paper, we propose a new method called LF-LPLC for multi-label learning, which integrates Label-specific features and local pairwise label correlation simultaneously. Firstly, we convert the original feature space to a low dimensional label-specific feature space, and therefore each label has a specific representation of its own. Then, we exploit the local correlation between each pair of labels by means of nearest neighbor techniques. According to the local correlation, the label-specific features of each label are expanded by uniting the related data from other label-specific features. With such a framework, it enriches the labels' semantic information and solves the imbalanced class-distribution problem. Finally, for each label, based on its label-specific features we construct a binary classification algorithm to test unlabeled instances. Comprehensive experiments are conducted on a collection of benchmark data sets. Comparison results with the state-of-the-art approaches validate the competitive performance of our proposed method.