Multi-label learning with missing labels using sparse global structure for label-specific features

Multi-label learning associates a given data instance with one or several class labels. A frequent problem with real life multi-label datasets is the lack of complete label information. Incomplete labels increase model complexity as the label correlation information is not reliable, resulting in a suboptimal multi-label classifier. Further, high dimensionality of multi-label datasets often introduces spurious feature-label dependencies. Thus, discovering label-specific features is imperative for efficient handling of high-dimensional data for multi-label learning with missing labels. To deal with the issues emerging from incomplete labels and high-dimensional input space, we propose a multi-label learning approach based on identifying the label-specific features and constraining them with a sparse global structure. The sparse structural constraint helps maintain the typical characteristics of the multi-label learning data. Instances are expressed as linear combination of label-specific features and the inter-relation guides the construction of model coefficients. The model also constructs supplementary label correlations to assist missing label recovery as part of the optimization problem. Empirical results on benchmark multi-label datasets highlight the effectiveness of the proposed method.