Scalable sparse bipartite graph factorization for multi-view clustering

Multi-view bipartite graph clustering (MBGC) has become an impressive branch of multi-view clustering (MVC) due to its remarkable scalability. Despite that various MBGC methods have been proposed, there are still some remaining issues. On the one hand, most of them need the singular value decomposition (SVD) of bipartite graphs to obtain spectral embedding, which may hampers efficiency when requiring a large number of anchors. On the other hand, the traditional sparsity-inducing norms like L₁ norm used in most methods fail to provide sufficient sparsity for embedding, which may impair effectiveness especially when facing noise and corruption. To this end, this paper proposes a scalable sparse bipartite graph factorization method for multi-view clustering (S²BGFMC). Specifically, to get rid of complex spectral analysis, the concept of bipartite graph factorization is proposed. In this concept, a more efficient partition technique, non-negative matrix factorization (NMF) is directly performed on bipartite graphs to maintain the efficiency of the whole clustering process. Additionally, L_2,log-(pseudo) norm, a column-wisely sparse, is constrained on the embeddings to bring the desired sparsity, thereby improving the effectiveness. To solve the proposed model, an efficient alternating iterative updating method is proposed. Extensive experiments illustrate that S²BGFMC can achieve superior efficiency and effectiveness against other baselines.