Multi-view feature extraction based on dual contrastive heads

Multi-view feature extraction can effectively address the problem of high dimensionality in multi-view data. Contrastive learning (CL), which is a popular self-supervised learning method, has recently attracted considerable attention. Most CL-based methods are only constructed at the sample level, which ignores the real structural information useful for feature extraction. In this study, we construct a structural-level contrastive loss, which promotes the consistency of the potential subspace structures in any two cross views in order to explore realistic and reliable structural information. On this basis, we propose a novel multi-view feature extraction method based on dual contrastive heads, which constructs the structural-level contrastive loss and integrates it into the sample-level CL-based method. In our method, the structural-level contrastive loss can help the sample-level contrastive loss extract discriminative features more effectively. Furthermore, the relationship between structural-level loss and mutual information, as well as the relationship between structural-level loss and probabilistic intra-class and inter-class scatter, are revealed, which provides the theoretical support for the excellent performance of our method. The final experiments on six real-world datasets demonstrate the superior performance of the proposed method compared to existing methods.