Research On Influence Maximization For Dynamic Social Networks

With the development of technology,the Internet has gradually become popular in human society,and people can form various online social networks through the Internet.In social networks,people can access and publish information more conveniently.The emergence of online social networks has made the speed and scope of information dissemination more extensive and rapid.In this situation,controlling the spread of information becomes crucial,as it can quickly and effectively disseminate knowledge,news,and other information,but it can also spread misunderstandings or inaccurate information quickly.In this context,people have proposed the influence maximization problem to optimize and control the spread of information.The influence maximization problem aims to find a user node set of size k,which can influence as many user nodes in the network as possible through a given influence propagation model.Currently,influence maximization has become one of the hottest research topics in the field of social computing and has been widely used in marketing,rumor control,political movements,and other fields.The development of online social networks has provided sufficient experimental data and rich application scenarios for influence maximization research,greatly promoting the development of this research direction.Currently,most influence maximization research is mainly carried out for static social networks with unchanged structures,while real social networks are constantly changing dynamically,and the research on the influence maximization of dynamic network is less.The dynamic characteristics of social networks can have a significant impact on influence or information diffusion,which cannot be ignored in influence maximization research.In response to the current research shortcomings,this article carries out a series of research on the dynamic social network influence maximization problem based on network representation learning and deep reinforcement learning,as follows:1.Most existing influence maximization research is directly based on network topology representation,but social networks are large in scale,and the high-dimensional properties of network topology representation can seriously affect the computational efficiency of algorithms.To address the above limitations,this article uses network representation learning methods to represent network nodes using low-dimensional vectors,attempts to solve the dynamic influence maximization problem in a low-dimensional dense latent vector space,and proposes a dynamic influence maximization algorithm based on network representation learning(DIMNRL)algorithm.The DIMNRL algorithm first constructs a limited node set for each node in descending order of similarity,then measures the influence size of each node by counting the number of times each node appears in these limited node sets,selects seed nodes,and uses incremental algorithm thinking to quickly update the seed set of networks at different times.Comprehensive experimental results on four real social network datasets,NetHEPT,Twitter,UCI,and Wikipedia,show that compared with the baseline method,the DIMNRL algorithm has similar performance in influence propagation,but the algorithm runs faster.2.The online influence maximization(OIM)problem aims to select the maximum influence seed node set from a social network with unknown activation probabilities.However,most existing online influence maximization research is aimed at static networks,where the activation probability of edges in the influence propagation model is fixed,and the relationship between network topology structure and influence propagation is not considered.In response to the above shortcomings,this article considers the correlation between edge activation probabilities and network structure,proposes the online dynamic influence maximization problem under dynamic social networks,and proposes a dynamic influence maximization algorithm based on deep reinforcement learning(DIM-DQN)to solve this problem.The DIM-DQN algorithm uses the DQN model to learn the strategy for selecting seed node sets,and selects seed nodes based on the influence propagation model and network structure,and updates the Q value function using the Bellman equation.Comprehensive experimental results on three real social,network datasets,including NetHEPT,Twitter,and UCI,show that the DIM-DQN algorithm outperforms several baseline methods in terms of influence propagation and computational efficiency.