A COLLABORATIVE DECISION-MAKING FRAMEWORK FOR INFLUENCE MAXIMIZATION BASED ON GRAPH CONTRASTIVE LEARNING AND DEEP REINFORCEMENT LEARNING

Abstract

To address the limitations of traditional influence maximization methods in complex social networks, including insufficient feature representation, strong dependence on predefined diffusion models, and limited capability for collaborative seed selection, this paper proposes a collaborative decision-making framework based on graph contrastive learning and deep reinforcement learning. First, the social network is modeled as a directed weighted graph, and a mathematical formulation of influence propagation and seed selection is established under the Independent Cascade(IC)model. Then, a graph contrastive learning strategy is introduced to conduct self-supervised pre-training over network topology, edge weights, and node attributes. By leveraging multi-view graph augmentation and node-level contrastive objectives, the framework learns high-quality node representations that better characterize diffusion potential. On this basis, the influence maximization task is reformulated as a sequential decision-making process, and a Double DQN-based seed selection mechanism is designed to iteratively select seed nodes. Marginal influence gain is adopted as the immediate reward to enable dynamic collaborative optimization of the seed set. Experiments conducted on six real-world social network datasets, including Petster-hamster, Tv-show, Politician, Advogato, Public, and Epinions, demonstrate that the proposed method consistently outperforms Random, PageRank, gIM, S2V-DQN, and ToupleGDD under different seed budgets. The results verify that combining the representation advantages of graph contrastive learning with the sequential decision-making capability of deep reinforcement learning can effectively improve influence maximization performance in complex networks.