GAWF: Influence maximization method based on graph attention weight fusion

Influence Maximization (IM) involves identifying a limited number of high-influence nodes within a network to maximize the number of influenced nodes. Although graph neural network-based IM methods have significantly improved generalization capability and propagation effect compared to traditional methods, they face challenges in capturing features of sparse graph networks and difficulties in computing gradient descent, leading to suboptimal diffusion effects and difficult model training. To address these issues, a Graph Attention Weight Fusion-based Influence Maximization method (GAWF) is proposed. First, the GAWF integrates adaptive weight decay optimization using AdamW with graph attention weight fusion. Second, inspired by the Beta-VAE in the CV field, an IM-VAE encoder method tailored for the IM problem is introduced. Finally, extensive comparative analysis experiments are conducted on four real datasets, including Jazz, Cora_ML and Power_Grid, to evaluate three traditional and four learning-based IM algorithms. Experimental results consistently show that the proposed GAWF method achieves a 0.1%[Formula: see text]6% improvement across various datasets, with more significant enhancements on sparser datasets, indicating that GAWF is reasonable and effective. Additionally, the proposed GAWF method holds promising applications in real-world scenarios such as social networks and public opinion analysis.