A Computationally Efficient Framework for Overlapping Community Detection in Large Bipartite Graphs

Abstract

Community detection, which uncovers closely connected vertex groups in networks, is vital for applications in social networks, recommendation systems, and beyond. Real-world networks often have bipartite structures (vertices in two disjoint sets with inter-set connections), creating unique challenges on specialized community detection methods. Biclique percolation community (BCPC) is widely used to detect cohesive structures in bipartite graphs. A biclique is a complete bipartite subgraph, and a BCPC forms when maximal bicliques connect via adjacency (sharing an (alpha, beta)-biclique). Yet, existing methods for BCPC detection suffer from high time complexity due to the potentially massive maximal biclique adjacency graph (MBAG). To tackle this, we propose a novel partial-BCPC based solution, whose key idea is to use partial-BCPC to reduce the size of the MBAG. A partial-BCPC is a subset of BCPC. Maximal bicliques belonging to the same partial-BCPC must also belong to the same BCPC. Therefore, these maximal bicliques can be grouped as a single vertex in the MBAG, significantly reducing the size of the MBAG. Furthermore, we move beyond the limitations of MBAG and propose a novel BCPC detection approach based on (alpha, beta)-biclique enumeration. This approach detects BCPC by enumerating all (alpha, beta)-bicliques and connecting maximal bicliques sharing the same (alpha, beta)-biclique, which is the condition for maximal bicliques to be adjacent. It also leverages partial-BCPC to significantly prune the enumeration space of (alpha, beta)-biclique. Experiments show that our methods outperform existing methods by nearly three orders of magnitude.