A Multi-instance Multi-label Learning Algorithm Based on Feature Selection

Multi-instance multi-label learning is an extension of multi-instance learning for multi-label classification. In order to select typical instances with high discrimination for multiple labels, the feature selection via Joint l(2,1) -norms minimization is introduced in this paper, and a multi-instance multi-label learning algorithm based on feature selection is proposed. All bags are mapped to typical instances after feature selection, and then the classifier considering label correlations is trained. Experimental results show that the proposed algorithm greatly improves the performance of multi-instance multi-label classifier compared with other methods.