Balanced and Discrete Multi-View Clustering With Adaptive Graph Learning

Graph-based methods have demonstrated strong performance in multi-view clustering (MVC) due to their capability to capture complex data structures. Among these, discrete spectral embedding learning has emerged as an effective strategy for directly producing clustering assignments, thereby avoiding potential suboptimality introduced by post-processing. However, most existing discrete MVC methods overlook the problem of skewed cluster assignments, which can significantly affect the quality and interpretability of clustering results in practical applications. To address this issue, we propose a novel framework for Balanced and Discrete Multi-view Clustering via Adaptive Graph Learning (BDMC-AGL). The proposed model jointly integrates adaptive graph construction and size-constrained spectral embedding learning into a unified optimization framework, enhancing the robustness of clustering while explicitly encouraging balanced partitioning. The introduction of size constraints into the discrete spectral embedding, however, poses a challenging optimization problem. To effectively solve it, we develop an efficient algorithm that guarantees convergence to an locally optimal solution. Extensive experiments conducted on benchmark datasets demonstrate that BDMC-AGL consistently outperforms state-of-the-art methods in terms of clustering accuracy and balance. Moreover, ablation studies validate the significant contribution of the size constraint mechanism in improving multi-view clustering performance.