FARGO: Feature-Augmented Representation Learning Combined With Enhanced GraphSAGE for Community Detection

Abstract

Community detection in complex networks remains a fundamental challenge across the social, biological, and technological domains. This paper presents FARGO, a graph neural network framework that unifies topological feature engineering, inductive learning, and community-aware optimization to address the limitations of existing methods in community detection. The framework enhances node representations by incorporating four structural centrality measures – Degree, clustering coefficient, PageRank and betweenness – before processing them through an enhanced GraphSAGE architecture. A novel composite loss function combines standard classification objectives with community balance regularization, thereby promoting structural cohesion through message passing rather than explicit constraints. Extensive experiments on seven real-world networks demonstrate the superiority of FARGO over the leading baselines, achieving consistent gains in accuracy and a significant improvement in Normalized Mutual Information (NMI). The framework preserves linear scalability and demonstrated substantially better performance in preserving minority communities and mitigating fragmentation issues that often arise in modularity-driven methods.