Efficient multi-view discrete co-clustering with learned graph

Graph-based multi-view clustering typically involves constructing view-specific similarity graphs, fusing graphs from multiple views, and performing two-step spectral clustering. However, several challenges arise: (1) constructing similarity graphs is computationally expensive, (2) balancing and integrating information across views is complex, and (3) the two-step clustering leads to information loss, solution deviations, and high computational cost. To address these issues, we present an efficient multi-view discrete co-clustering framework. It automatically learns a multi-view consistent anchor similarity matrix, dynamically weighting the contributions of different views based on the original data structure and evolving indicators. The resulting anchor similarity matrix serves as the weight matrix for a bipartite graph, facilitating efficient co-clustering of raw data and anchors. Additionally, we introduce a time-economical optimization algorithm to solve for discrete indicators directly. Extensive experiments demonstrate that the proposed method outperforms multiple competitors, highlighting its superior performance and efficiency. The code is available at https://github.com/caccode/EMDC.