Multi Resolution based Graph Convolutional Networks for Detecting Communities in Social Networks

Detecting the fundamental community structure of a social network is crucial in social network analysis. All though, there are many recent approaches in detecting communities, analysing large scale social networks at all resolutions is a challenging task. To solve this problem, the proposed research focuses on developing an efficient Multi Resolution-Graph Convolutional Network (MR-GCN) to identify the communities in social networks. The proposed method with the graph mining technique improves the output graph quality in identifying the communities. The conventional approaches depend on the labelled data, which is limited data in the real time applications. Initially, the nodes of raw graph are converted into embeddings and are further processed by multiple layers of a GCN to incorporate both the local and global structural features. The MR feature fusion is used to combine node embeddings from different resolutions, which includes all necessary information. The final phase includes Adam optimizer to train the model, and the overall experimental results obtained are accuracy of 97.85%, precision of 96.57%, and f1-score of 95.77%. The proposed method results are compared with the existing methods such as Graph Neural Network encoding scheme for a Multi-Objective Evolutionary Algorithm (GNN-MOEA) and Markov Random Fields GCN (MRFasGCN).