Clustering with Dynamic Bipartite Graph Learning

Abstract

Anchor-based bipartite graph clustering algorithms greatly enhance data analysis by accelerating computations without sacrificing performance. However, these methods face two limitations: reliance on fixed bipartite graphs and lack of interaction between the bipartite graph and anchor labels, which reduces adaptability to dynamic data and clustering performance. To overcome these, a novel method, called Clustering with Dynamic Bipartite Graph Learning (DBGL) is proposed. The DBGL objective function is composed of two parts: construct the bipartite graph and learn the anchor labels. The bipartite graph and anchor labels are updated iteratively, resulting in a dynamic bipartite graph that evolves throughout the clustering process. Our dynamic technique is more flexible and performs better on large, complicated datasets than fixed bipartite graph methods. Additionally, we introduce label transmission into DBGL, enabling autonomous interaction between bipartite graph and anchor labels, which mutually guide each other toward optimal clustering results. We design an alternating optimization algorithm where the bipartite graph is updated using an Iteratively Re-Weighted (IRW) algorithm, while the anchor labels are optimized through an improved Coordinate Descent (CD) algorithm. Comprehensive experiments on benchmark datasets demonstrate that DBGL outperforms leading methods in clustering accuracy, efficiency, and robustness, with theoretical analysis confirming its convergence and stability.

Keywords: Clustering, dynamic bipartite graph, Optimization, label transmission