DGN: influence maximization based on deep reinforcement learning

DGN: influence maximization based on deep reinforcement learning

연구 분야: Artificial Intelligence

학회: The Journal of Supercomputing

The quality of seeds is crucial for influence maximization in social networks. Seed selection algorithms based on deep reinforcement learning (DRL) combine the representation capabilities of deep learning with the decision-making abilities of reinforcement learning, showing great potential in solving the influence maximization problem. In this paper, an end-to-end trained dual coupled graph neural network (DGN) for seed selection is proposed. This method not only ingeniously utilizes dual coupled graph neural networks for node embedding but also employs a deep Q-network (DQN) from deep reinforcement learning to explore the topology and node attribute information of social networks to generate context-rich node vector representations. Moreover, it can train the network to approximate the Q-function through DQN without prior knowledge, allowing it to adaptively select the optimal strategy based on the current state and action, thereby maximizing the influence spread of the selected seed nodes. Extensive experiments on synthetic and real networks demonstrate that the proposed DGN not only achieves performance very close to or even surpassing current state-of-the-art models like ToupleGDD but also exhibits better robustness, making it more suitable for the influence maximization problem in today’s large-scale, complex, and dense social networks.