Robust Trust Optimization in Social Graphs: Centrality-Aware GNNs with Genetic Optimization

Abstract

Precisely evaluating trust inside social networks is essential for recognizing reliable users amid chaotic and varied interactions. This research introduces a triadic hybrid model that integrates structural centrality measures, a Centrality-Aware Graph Attention Network (CA-GAT), and Genetic Algorithm (GA) optimization. At the outset, centrality metrics degree, closeness, betweenness, eigenvector, and PageRank are standardized and integrated to produce a comprehensible initial structural trust measure. Upon application to a processed graph derived from Epinions data, which consists of 4,039 nodes and 76,879 edges, the most robust individual centrality baseline attained an ROC-AUC of 0.67. The CA-GAT model improved the structural depiction and elevated the ROC-AUC to 0.83, thereby pinpointing 3,360 nodes that exceeded the trust threshold. Following this, a GA-driven optimization phase further refined the final selection of reliable nodes, identifying 3,005 nodes and elevating the ROC-AUC to 0.85. Ablation ana