Multi-view representation learning with dual-label collaborative guidance

Abstract

Multi-view Representation Learning (MRL) has recently attracted widespread attention because it can integrate information from diverse data sources to achieve better performance. However, existing MRL methods still have two issues: (1) They typically perform various consistency objectives within the feature space, which might discard complementary information contained in each view. (2) Some methods only focus on handling inter-view relationships while ignoring inter-sample relationships that are also valuable for downstream tasks. To address these issues, we propose a novel Multi-view representation learning method with Dual-label Collaborative Guidance (MDCG). Specifically, we fully excavate and utilize valuable semantic and graph information hidden in multi-view data to collaboratively guide the learning process of MRL. By learning consistent semantic labels from distinct views, our method enhances intrinsic connections across views while preserving view-specific information, which contributes to learning the consistent and complementary unified representation. Moreover, we integrate similarity matrices of multiple views to construct graph labels that indicate inter-sample relationships. With the idea of self-supervised contrastive learning, graph structure information implied in graph labels is effectively captured by the unified representation, thus enhancing its discriminability. Extensive experiments on diverse real-world datasets demonstrate the effectiveness and superiority of MDCG compared with nine state-of-the-art methods. Our code will be available at https://github.com/Bin1Chen/MDCG.