Community Detection in Social Networks: the Impact of Clustering on Graph-Based Deep Learning - a Study

Social network(SN) represents the structure of individuals connected by relationships such as friendships or collaborations. Social Network Analysis (SNA) provides a complete understanding of the topological structures, semantic relationships, and their evolutions. One major area of research in SNA is the detection of communities. These communities represent different social or semantic structures that can be achieved by several methods, including Graph-Based Models, Modularity-Based Probabilistic Models, Label Propagation, Clique Percolation, Clustering, Random Walks, and Deep Learning (DL) Techniques. Despite their individual benefits, there are problems with scalability, interpretability, and precise community identification in large or complex networks. There is uncertainty over the utility of clustering techniques for Community Detection (CD) because few works employ them, and few do not. The aim of this study is to analyse the performance of Graph-based DL methods in CD with respect to their performance when combined with the clustering algorithms. Graph-based DL methods are ideal for CD because they leverage the inherent structure of networks. Models such as Graph Attention Networks (GAT), Graph Convolutional Networks (GCN), GraphSAGE, and Graph Transformer Networks (GTN) are independently applied to the CORA dataset in order to extract node embeddings to identify communities. Following this, the K-means and Hierarchical clustering algorithms are integrated with the aforementioned methods for the CD task. Results indicate that CD methods without the incorporation of clustering algorithms generally yield superior results than with clustering.