Research On Information Dissemination Control Based On Reinforcement Learning

Online social network(OSN)has been a very important carrier of information dissemination in contemporary society,and information dissemination is the core function in the field of online social network,and also the main research content.When information is disseminated in the network,some undesirable information is easily spread in the network,because these undesirable information may be contrary to the public ethics,moral concepts,laws and regulations,and it is likely to mislead people and even cause economic and property losses.Therefore,proper immune control of information in social networks is needed in order to reduce the bad effects caused by bad information.The current study of information dissemination is of great practical significance for many fields such as public opinion monitoring,and also has quite broad application prospects.In this dissertation,we first analyze the shortcomings inherent in previous preventive immunization strategies,which make the previously defined immunization strategies ineffective in controlling information dissemination,or use a great cost to suppress the dissemination of information.In this dissertation,we address this shortcoming by immunizing a small population to minimize the cost of immunization control and using reinforcement learning methods to achieve different immunization strategies for different information propagation states.This is specifically achieved with a data initialization process,an information propagation process and a policy generation process that minimizes the cost of immunization.First,in the data initialization process,the nodes,edges,and weights in the datasets are transformed into the form of adjacency matrix after normalization.Then,in the information propagation process,an independent cascade model is used to simulate the random propagation pattern of information in real life,and the initial nodes of information propagation are used as input,and the model will output the list of activated nodes.Finally,in the policy generation process of minimizing the immunization cost,the policy network and the evaluation network are combined,the policy network outputs the actions,the evaluation network scores each action,and the results of the scores in turn guide the actions executed by the policy network,and this process optimizes the immunization control policy and finally outputs the minimum immunization cost.Simulations and comparisons were conducted on 4data sets to evaluate the performance of the information dissemination control policy algorithm proposed in this dissertation,and the experimental results show that this dissertation’s policy outperforms existing information dissemination control policies in the field of information dissemination control.This dissertation also improves the immune policy generation algorithm with Monte Carlo tree search and deep residual network algorithm based on reinforcement learning algorithm.The role of Monte Carlo tree search method is to shorten the search speed of neural network,and the role of deep residual network method is to reduce the neural network error.The immune strategy framework is improved by adding a new deep residual network before the strategy network and the evaluation network,and a new Monte Carlo tree search method after the strategy network and the evaluation network,so that the Monte Carlo tree search and the neural network are combined,and the neural network parameters are optimized by the Monte Carlo tree search method,which in turn guides the Monte Carlo tree search by the optimized neural network,and this process optimizes the strategy network and evaluation networks.Experimental results show that the incorporated Monte Carlo tree search and deep residual network models are less costly to immunize with the reinforcement learning approach and have a smaller immunization cost compared to the two strategies.