Cross-view alignment guided deep anchor multi-view clustering

Multi-view clustering (MVC) aims to uncover intrinsic data structures by leveraging both the consistency and complementarity among multiple views. While anchor graph clustering enhances scalability by approximating sample-wise graphs using a small set of representative anchors, existing deep anchor-based MVC methods typically select centroid-like anchors, which may fail to capture high-density semantic regions in high-dimensional spaces. Furthermore, anchor graphs are often constructed independently for each view, leading to inconsistent local topologies and misaligned global semantics across views. In addition, enforcing cross-view consistency directly in the feature space may interfere with reconstruction-driven embedding learning, creating a trade-off between clustering agreement and representation fidelity. To address these challenges, we propose CADA-MVC, a cross-view alignment-guided deep multi-view anchor clustering framework. CADA-MVC employs a hierarchical anchor selection strategy guided by local density peaks in the latent space, aligns view-specific anchor structures with a consensus structure through soft matching, and promotes cross-view agreement via a soft-threshold contrastive objective in the assignment space—without imposing direct constraints on the learned embeddings. Experiments on six benchmark datasets demonstrate that CADA-MVC outperforms most representative baseline methods while preserving the computational efficiency of anchor graph approximation.