Topic-aware influence maximization based on deep reinforcement learning and group relative optimization

Information diffusion on social networks is increasingly complex and diverse. Identifying suitable users to recommend and spread specific information is critical for social network analysis. The Topic-aware Influence Maximization (TIM) problem aims to identify a seed set that maximizes the influence spread under a given topic distribution. However, existing TIM methods suffer from severe computational inefficiencies. Meanwhile, current Deep Reinforcement Learning (DRL)-based methods mostly ignore the interplay between network structure and topic heterogeneity. To address these challenges, this paper proposes GR-TIM, an end-to-end DRL-based framework for TIM. GR-TIM first estimates the pre-global influence using Graph Neural Networks (GNNs). Then, a group relative optimization strategy partitions users based on topic and community structures. We further leverage intra-group collaboration to apply the global-local optimization paradigm to agent training and inter-group competition to achieve adaptive seed selection for the target topic. Experiments on six real-world datasets demonstrate that GR-TIM outperforms state-of-the-art DRL-based methods in terms of multi-topic influence spread and reduces runtime by two to three orders of magnitude compared to existing simulation-based methods.