A model framework for the enhancement of community detection in complex networks

Community detection is an important data analysis problem in many different areas, and how to enhance the quality of community detection in complicated real applications is still a challenge. Current community detection enhancement methods often take the enhancement as a preprocess of community detection. They mainly focus on how to design the suitable topological similarity of nodes to adjust the original network, but did not consider how to make use of this topological similarity more effectively. In order to better utilize the similarity information, we propose a model framework which integrates the enhancement into the whole community detection procedure. First, we calculate the structural similarity of nodes based on network topology. Second, we present a stochastic model to describe the community memberships of nodes; we then model the strong constraint based on structural similarity, i.e., we make each node have the same community membership distribution with its most similar neighbors; and then we model the weak constraint, i.e., if two nodes have a high similarity we will make their community membership distributions close, otherwise we will make them not close. Finally, we present a nonnegative matrix factorization approach to learn the model parameters. We evaluate our method on both synthetic and real-world networks with ground-truths, and compare it with five comparable methods. The experimental results demonstrate the superior performance of our new method over the competing ones for community detection and enhancement.