Proxy-Based Dynamic View Alignment With Cross-View Structure Preservation for Multi-View Clustering

Multi-view clustering aims to fully discover consistent clustering structures across different views, thereby improving the clustering performance. However, they usually lose structural information between different views during multi-view feature integration. Additionally, the misalignment of clustering centers from different views may lead to performance degradation. To tackle these challenges, in this paper, we propose a novel approach, called proxy-based dynamic view alignment with cross-view structure preservation (PDVA-CSP), for multi-view clustering. First, we design a graph aggregation module to explore the structure information between views through graph-based aggregation. Then we propose a proxy-based dynamic alignment strategy based on attention mechanisms to address the misalignment of proxies across views. Finally, we integrate them into an end-to-end learning framework, and then optimize it via a joint reconstruction loss and contrastive learning framework, seamlessly integrating feature extraction, view alignment, and clustering. The experimental results on several multi-view datasets demonstrate that the proposed PDVA-CSP method significantly outperforms other state-of-the-art methods in multi-view clustering tasks. The source code for this work will be available later.