Incorporating view location information for multi-view multi-label learning

Abstract

In multi-view multi-label learning (MVML), subspace learning provides an effective solution for integrating multi-view data. However, the current learning concentrates on the shared subspace construction process and neglects the exploration of location information, i.e., the view source of the feature, which is inherent to the feature itself. This has somewhat limited the progress of research on MVML. On this basis, we propose a multi-view multi-label learning method based on incorporating view location information (MMVLI) for the first time with reference to the Vision Transformer architecture. Firstly, the common features and special features between views are obtained through the feature extraction. The above two are combined according to the feature dimension and added into the learnable location information encoding matrix named the collaborative matrix. Then, the collaborative matrix is put into Transformer Encoder for self-attention learning to get the final feature matrix for multi-label learning. Extensive comparative experiments with existing state-of-the-art methods on six publicly available multi-view multi-label datasets demonstrate the effectiveness of MMVLI.