Community Detection through Recursive Partitioning in Bayesian Framework

Community detection involves grouping the nodes in the network and is one of the most-studied tasks in network science. Conventional methods usually require the specification of the number of communities $K$ in the network. This number is determined heuristically or by certain model selection criteria. In practice, different model selection criteria yield different values of $K$, leading to different results. We propose a community detection method based on recursive partitioning within the Bayesian framework. The method is compatible with a wide range of existing model-based community detection frameworks. In particular, our method does not require pre-specification of the number of communities and can capture the hierarchical structure of the network. We establish the theoretical guarantee of consistency under the stochastic block model and demonstrate the effectiveness of our method through simulations using different models that cover a broad range of scenarios. We apply our method to the California Department of Healthcare Access and Information (HCAI) data, including all Emergency Department (ED) and hospital discharges from 342 hospitals to identify regional hospital clusters.