Feature selection for multi-label learning with missing labels

In multi-label learning, feature selection is a non-ignorable preprocessing step which can alleviate the negative effect of high-dimensionality. To address this problem, a number of effective information theory based feature selection algorithms for multi-label learning are proposed. However, these existing algorithms assume that the label space of multi-label training data is complete. In practice, the standpoint does not always hold true, due to the ambiguity among class labels or the cost effort to fully annotate instances. In this paper, we first define the new concepts of multi-label information entropy and multi-label mutual information. Then, feature redundancy, feature independence, and feature interaction are defined, respectively. In which, feature interaction is used to select more valuable features which may be ignored due to the incomplete label space. Moreover, a multi-label feature selection method with missing labels is proposed. Finally, extensive experiments conducted on eight publicly available data sets verify the effectiveness of the proposed algorithm via comparing it with state-of-the-art methods.