Data driven framework for ranking influential nodes in social networks using machine learning and neural networks

Abstract In the dynamic and intricate domain of social networks, precise identification and ranking of influential nodes are essential for effective information dissemination, strategic decision making, and optimal resource allocation. Conventional centrality metrics frequently yield unstable outcomes due to their limited capacity to capture the multidimensional nature of influence. To address this challenge, a novel hybrid data-driven framework is introduced, comprising three integrated stages. Initially, sophisticated machine learning algorithms, namely Random Forest, XGBoost, and CatBoost, are applied to determine objective weights for six centrality metrics: The following centrality measures are examined: degree centrality, betweenness centrality, closeness centrality, eigenvector centrality, PageRank, and the global clustering coefficient-dependent degree centrality. These weights are derived from real-world data, guided by the Susceptible Infected Recovered diffusion model. In the subsequent stage, the TOPSIS method consolidates these weighted metrics to generate preliminary node rankings, thereby establishing training labels for subsequent modeling. In the final stage, a Multi Layer Perceptron neural network predicts influence scores with enhanced autonomy and precision. This network leverages structural features to rank nodes without requiring repeated TOPSIS execution. The validation of the proposed framework was conducted on four benchmark datasets: Zachary's Karate Club, American College Football, Dolphin Social Network and Bitcoin-OTC. The findings indicate that the proposed framework demonstrates superior accuracy, robustness to structural noise, and generalizability across both unsigned and signed graphs when compared to traditional single metric approaches. The outcomes underscore the framework's adaptability across diverse network structures, offering a robust instrument for advancing network science research and applications.