Multi-Label Learning with Missing Features

Multi-label learning deals with the problem that each example is associated with multiple class labels simultaneously. Existing multi-label learning approaches all assume the feature space is completed and construct classification models using examples with sufficient feature information. However, in real-world applications, it is difficult to get a fully completed feature matrix, that is, only partial feature information of each example can be obtained. In this paper, we formalize this problem as multi-label learning with missing features. To tackle this problem, we learn a feature correlation matrix and apply it to obtain a new supplementary feature matrix, which has richer feature information than the original missing feature matrix. After that, to improve the performance of multi-label classification, we constrain feature correlation on coefficient matrix by assuming that if two features are strongly correlated, the similarity between their corresponding parameter vector will be large. Besides, we also constrain label correlation on output of labels to capture more sufficient relationships between different labels. Extensive experiments show a competitive performance of our method against other state-of the-art multi-label learning approaches.