Method For Influence Maximization Of Unknown Social Network Based On Reinforcement Learning

With the development of communication technology and mobile Internet,the research on the influence communication of social networks has become a hot research issue.Among them,the influence maximization problem refers to selecting a certain number of seed nodes under a given influence propagation model to spread the information to the widest range.The application scenarios of influence maximization are very wide,ranging from recommendation system and virus marketing to information diffusion and link prediction.The existing influence maximization methods have the problems of large amount of calculation,slow calculation speed and difficult to migrate to unknown social networks.Aiming at the problems existing in the existing influence maximization algorithms on social networks with unknown network structure,this paper investigates and studies,models the influence maximization problem of unknown social networks by reinforcement learning,and proposes a method to maximize the influence of unknown social networks based on reinforcement learning.Specifically,the method models the influence maximization problem on the social network with unknown network structure as a multi-agent cooperation problem,and uses exploration agent and selection agent to explore the network structure and select seed nodes respectively.According to the special structure of social network,a reinforcement learning algorithm adapted to graph structure is proposed based on DQN and QMIX algorithm,and a solution to the problem of reward allocation in this scenario is proposed.At the same time,the method makes use of the special properties of social networks,transforms the exploration strategy of the algorithm,and improves the learning ability of the algorithm.In this paper,experiments are carried out in unknown social networks of various scales,and some existing heuristic methods are compared.The results show that the algorithm proposed in this paper has achieved good performance compared with the classical algorithms in this field in terms of final influence propagation effect,calculation speed and generalization.