Graph Attention Network for Interpretable Graph Community Detection

Identifying communities within complex networks is crucial for understanding their organization and function. While numerous community detection algorithms exist, many lack inherent explainability, hindering practical application and trust. To address this, we propose IGCD, an Interpretable Graph Community Detection framework. Our approach first utilizes the Louvain algorithm for an initial partitioning of the network. Subsequently, a Community-aware Graph Attention Network (GAT) refines these initial assignments, leveraging node features and network structure. Finally, gradient-based techniques are employed to provide explanations for the resulting community assignments by quantifying the importance of input features. Experiments on a real-world social network demonstrate that our method achieves competitive accuracy in community detection while offering valuable insights into the factors driving community formation.