Detecting communities in attributed networks through bi-direction penalized clustering and its application

Abstract

Exploiting heterogeneous information in attributed networks to improve the performance of community detection has attracted considerable research attention. Although variational graph autoencoder (VGAE)-based methods have been proven to be effective strategies, they perform community detection based on assumptions regarding the dimension of embedding and the number of communities, limiting their effectiveness and applicability. In this study, we combined VGAE-based methods and a bi-direction penalized clustering algorithm (BiPClust) for community detection. Our approach addresses the issues of dimension selection and community number determination by automatically optimizing penalized clustering. Both the computational algorithm and statistical theorems confirm that BiPClust effectively mitigates the impacts of redundant embedding and determines the unknown number of communities. Furthermore, applying the proposed methods to community detection on benchmark datasets and syndicated investment networks in China reveals that BiPClust surpasses other methods in performance.