Consensus-guided individual graph learning via enhanced tensor low-rank for robust multi-view clustering.

Most graph-based multi-view clustering methods first learn individual similarity graphs for each view and then derive a consensus graph from these similarity graphs. However, these methods suffer from several limitations, such as irreversible information loss caused by constructing similarity graphs from high-dimensional noisy data, error accumulation in the consensus graph due to iterative optimization, and the oversimplification of inter-view relationships that overlooks the synergy among diversity, consistency, and higher-order correlations. To address these issues, we propose a novel method named Consensus-guided Individual Graph Learning via Enhanced Tensor Low-Rank (CIGETL). Unlike existing methods, CIGETL uses a consensus graph to guide the learning of individual graphs, improving consistency and capturing shared information across views. Specifically, CIGETL first learns consistent representations in a low-dimensional common subspace and constructs a consensus graph. The consensus graph is then reconstructed in each view, serving as a dictionary for self-representation learning. Double Laplacian manifold constraints balance diversity and consistency across views, while column sum constraints enhance adaptability. Finally, CIGETL uses an enhanced tensor low-rank minimization method to capture higher-order correlations between views. Extensive experiments on six public datasets show that CIGETL outperforms existing multi-view methods in clustering performance and robustness.