Correntropy-Based Bipartite Graph Factorization for Clustering

Shangzong Yang, Ben Yang, Jinghan Wu, Zhiyuan Xue, Xuetao Zhang

Abstract: Non-negative matrix factorization (NMF) is widely utilized in the domain of clustering, primarily due to its efficacy in decomposing the initial matrix into two smaller matrices, thereby facilitating the discernment of underlying data characteristics. Nevertheless, existing NMF-based methods still face two critical challenges: 1) the clustering efficiency is significantly affected by the original matrix’s dimensionality. 2) in the presence of nonlinear and non-Gaussian noise and outliers, their robustness markedly declines. To tackle these issues, we propose a correntropy-based bipartite graph factorization model for clustering (CBGFC). First, a bipartite graph is constructed to capture the structure of samples, providing a more suitable representation. Then, by integrating the bipartite graph and NMF into a unified clustering framework, we avoid the efficiency being affected by the dimensionality of the data. Additionally, to improve the robustness of CBGFC, correntropy is introduced into the clustering model to handle noise and outliers. Extensive experiments demonstrate that CBGFC outperforms other state-of-the-art baselines in terms of clustering efficiency and robustness.