Incomplete Multi-view Clustering via Cross-view Diversity Enhanced Fine-grained Tensor Completion

Despite the significant attention that incomplete multi-view clustering (IMC) has attracted for analyzing incomplete data, most existing methods fail to capture intricate cross order and cross-view latent relations or alleviate the adverse impact of severe missingness. In this paper, we propose a novel approach to address these limitations. Specifically, we construct a fine-grained tensor via cross-order neighbor structures for tensor completion, which provides a high-quality initial graph for the downstream model. We then employ log-based tensor rank approximation to capture latent structural information from the handcrafted graph tensor. Meanwhile, a fusion strategy with self-adaptive weighting is designed to preserve cross-view consistency and diversity in the completed tensor. All modules are integrated into a unified framework, and extensive experiments verify the effectiveness and superiority of our method under various scenarios and missing data conditions.