Research on node importance in complex networks continues to draw on a wide range of approaches, including classical centrality metrics, entropy-based measures, and various machine-learning models. Previous studies often reports that entropy-driven and learning-based techniques yield higher accuracy and faster convergence for tasks such as information spreading or robustness analysis, although these results can vary depending on the dataset and the assumptions of each method. What remains less explored is the extent to which these techniques account for the contextual or semantic characteristics of nodes, which are not directly visible in structural data. In this study, a hybrid framework is introduced to address this gap by combining established centrality metrics with evaluations produced by a large language model (LLM). The approach was examined on several well-known benchmark networks—Zachary’s Karate Club, the Krackhardt Kite Graph, and the Jazz Musicians Network—and the resulting rankings showed clearer separation among nodes and, in many cases, more interpretable patterns than the purely structural baselines. To situate these findings within the broader literature, results were compared with three prior studies that employ composite or entropy-based centrality formulations. To explore scalability, three larger collaboration networks from the SNAP repository were also evaluated. The overall evidence suggests that LLM-supported scoring may offer a complementary perspective for complex network analysis, especially when structural information alone provides an incomplete picture.
Sabah et al. (Wed,) studied this question.