A residual graph reinforcement learning for budgeted influence maximization

Abstract

The Budgeted Influence Maximization (BIM) problem has been widely applied like critical node identification, social media marketing, and rumor suppression. Nowadays, the networks exhibit significant variations in both structure and scale. Graph Reinforcement Learning (GRL)-based methods are commonly employed to solve the BIM problems. However, the GRL-based methods suffer from lower performance in spread effect with diverse structures and scales. This limitation arises from their inability to effectively capture and leverage the intricate relationships within local neighborhood structures. To address this, we propose a novel Residual Graph Reinforcement Learning (RGRL)-based framework, which incorporates three key designs: (1) A network partitioning mechanism based on community detection algorithms to identify dense subgraphs; (2) A hybrid framework combining global training with subgraph execution to enhance the search capability of RGRL, particularly for effectively learning both subgraph-level and global neighborhood information; and (3) A residual graph neural network that captures global information and subgraphs neighborhood information for better selecting the seed node. We evaluate our proposed RGRL model alongside five baseline methods on four open-source datasets of varying sizes. The experimental results demonstrate that RGRL achieves the maximum improvements in IC, LT, and SIS information propagation models, with increases of up to 11.21%, 60.80%, and 12.91%, respectively.