Fine-grained tensor completion for incomplete multi-view clustering

Incomplete multi-view clustering (IMC) is to cluster data from multiple views, where some views may contain missing entries. In this paper, we propose a novel method for IMC. First, we construct cross-order neighbor graphs of the observed entries in different views, which preserve comprehensive and complementary information of the data. Then, we recover a fine-grained low-rank tensor of the incomplete data from the incomplete cross-order neighbor graphs to preserve cross-view consistency using the low-rank tensor completion technique, where a novel log-based tensor approximation is developed for rank constraint. Moreover, we incorporate clustering-driven structural constraints with auto-adjusted weights to preserve cross-view diversity, which promotes the recovered low-rank tensor to be more suitable for clustering applications. Extensive experimental results confirm the superiority of the proposed method compared with state-of-the-art IMC methods.