Abstract:  In order to effectively utiliz the multi-view data information and enhance the feature selection performance, the structured multi-view sparse regularization was constructed, based on which a novel supervised feature selection method, namely structured multi-view supervised feature selection (SMSFS), was proposed. SMSFS could simultaneously consider the importance of each view features and the importance of individual feature in each view to combine the multi-view data information effectively in the feature selection process, and then to boost the supervised feature selection performance. Because the objective function of SMSFS is non-convex, an effective iterative algorithm was proposed to solve the objective function. The proposed structured multi-view supervised feature selection method SMSFS was applied into image annotation task and extensive experiments were performed on NUS-WIDE and MSRA-MM2.0 image datasets. The proposed method SMSFS was compared with other feature selection methods and the experimental results demonstrated the effectiveness of SMSFS, which means that it could effectively utilize the multi-view data information to boost the feature selection performance.

Key words: multi-view learning, structured sparse regularization, supervised feature selection