Research Of Influence Maximization And Information Cascade Prediction Based On Representation Learning

With the rapid development of internet social platforms,more and more behavior log data are generated on social platforms.How to effectively mine and utilize valuable information from these data is of great significance to social network analysts.At the same time,social network influence maximization and information cascade prediction have also become the research focus of researchers in social network analysis.he purpose of social network influence maximization is to obtain users with high influence or to mine the user set with the widest range of information influence.In view of the deviation between the simulated information diffusion model and the reality,this paper proposes to solve the influence maximization problem by using the cascade data and mining the active forwarders in the cascade.For the problem of information cascade prediction,the cascade increment or cascade scale of information is usually predicted based on user characteristics.In view of the existing research on information cascade prediction,the prediction effect of using artificial features or social graph structure features is not ideal.Meanwhile,so as to solve the problem of feature redundancy affecting downstream prediction tasks,this paper proposes to solve the information cascade prediction by using the feature model of graph embedding combined with PCA.The details of work are as follows:1)Most of the existing researches on social network influence maximization are based on the network graph structure,which does not utilize the information contained in the cascaded data and cannot effectively capture the real influence among users.To solve this problem,this paper proposes a method of influence maximization based on cascade data fusion cascade active forwarding.Firstly,an embedded neural network model integrating active forwarders was designed,and the user feature vector was obtained through the supervised training of cascade forwarding records.Then,the user influence of the integrated active forwarders was calculated according to the number of information reachable objects and diffusion probability.Finally,the seed set was selected based on greedy strategy.Compared with four representative methods on three large-scale data sets,the experimental results show that the proposed method is more effective in the real diffusion range of information.2)In view of the fact that graph embedding algorithm can effectively excavate the hidden features of nodes,and to avoid the redundancy of graph embedding features,this paper proposes an information cascade prediction method based on graph embedding combined with PCA.Firstly,this method obtains the feature vector of information publisher by embedding the cascade graph,then uses PCA to embed the feature principal component analysis into the graph,trains the machine learning regressor by using the principal component feature,and finally uses the regressor to predict the cascade increment of information.The experimental results show that this method can improve the effect of information cascade prediction by combining graph embedding and regressor with PCA.