Gumbel-SoftMax based graph convolution network approach for community detection

Abstract

Lately, there has been exploration and successful implementation of various approaches of deep learning in different systems such as visual processing, pattern recognition, speech recognition, etc. In this research work, we explore the possibilities of employing the deep learning methods like Graph Neural Network (GNN) to unveil hidden communities among the complex networks. Complex systems have gained mounting attention in numerous areas, including biological networks, social networks, the recommender system, information graphs, and also life sciences. Scientific upgradation pertinent to network analysis is enabled by the role of deep learning models in forming an interaction of nodes in a graph. In the research paper, we propose a deep learning model, Gumbel-SoftMax based Graph Convolution Network (GS-GCN) approach to discern the clusters present in complex networks. We have also used the Gumbel-SoftMax measure to retrieve the relevant features of the given networks. We have used two-layer graph convolution network model. In this model, we have utilized two parameters, degree matrix and adjacency matrix with loops to enhance the performance. The experimental outcomes substantiate that the proposed approach outperforms the conventional methods noticeably, which clearly demonstrates the efficacy of deep learning models in uncovering the communities hidden in graph networks. We have performed rounds of testing on the proposed GS-GCN method, using four real-world datasets, namely, Zachary Karate club, Dolphins, Political Books, and Les Misérables. The fact that the proposed method discovers better community structures is proved by the results of the detailed experiments.