Toward embedding-based multi-label feature selection with label and feature collaboration

Similar to single-label learning, multi-label learning employs feature selection technique to alleviate the curse of dimensionality. Many multi-label methods, which utilize label correlation or instance correlation to select meaningful features, were proposed in recent years. However, these multi-label feature selection methods explored the label correlation or instance correlation via similarity measures, which may not perform well in revealing complex relationships between labels and instances. Furthermore, label correlation and instance correlation are employed as independent strategy to select the discriminative features, and no general framework can currently be considered the two together as to their effect. In this paper, we propose a new multi-label feature selection method named CMFSS, which explicitly explores the label correlation and instance correlation in a collaborative manner. Firstly, CMFSS learns the label correlation and the instance correlation via the ADMM technique. Secondly, the learned label correlation and instance correlation are seamlessly incorporated into the multi-label feature selection model. Finally, CMFSS utilizes l(2, 1)-norm as sparsity regularization to control the model complexity. Extensive empirical evaluations conducted on multiple benchmark datasets clearly show the superiority of the proposed multi-label feature selection method.