Abstract: To address the issue of degraded propagation performance of seed node sets in large-scale social networks by existing deep reinforcement learning models, this study proposes an end-to-end deep reinforcement learning model called EDG-DQN. First, it em⁃
ploys an attention mechanism to focus on key information, constructing a dual-channel graph attention network to learn embedded representations of nodes in both the influence application space and the influence reception space. Second, a variant of the long short-term memory network structure——the dynamic influence unit is introduced, treating the environmental state in reinforcement learning as a temporal feature to dynamically enhance the spatial embedded representations of nodes. Finally, the deep Q-network approximates the Q-function to achieve an optimal selection strategy for nodes. Comparative experiments based on dynamic simulated network propagation and simulated real-world network computational costs demonstrate that the seed node sets selected by this model on large-scale sparse social network datasets exhibit superior influence and diffusion efficiency compared to existing benchmark models.

Key words: influence maximization, deep reinforcement learning, large-scale social networks