Entropy-Based Active Learning for Precise Influence Evaluation in Complex Networks

Evaluating node influence is fundamental for identifying key nodes in complex networks. Existing methods typically rely on generic indicators to rank node influence across diverse networks, thereby ignoring the individualized features of each network itself. Actually, node influence stems not only from general features but also multiscale individualized information encompassing specific network structure and task. Here, we design an active learning (AL) architecture to predict node influence quantitatively and precisely, which samples representative nodes based on graph entropy correlation matrix integrating multiscale individualized information. This brings two intuitive advantages: 1) discovering potential high-influence but weak-connected nodes and 2) improving the influence maximization (IM) strategy by deducing influence interference. Significantly, our architecture demonstrates exceptional transfer learning capabilities across multiple types of networks, which can identify those key nodes with large disputation across different existing methods. In addition, our approach, combined with a simple greedy algorithm, exhibits dominant performance in solving the IM problem (IMP). This architecture holds great potential for applications in graph mining and prediction tasks.