Research On A Propagation Model Based On Representation Learning

Social network is the third wave of Internet development after portal website and search engine.Since 2000,social platforms such as facebook and twitter have sprung up one after another,and various services are changing with each passing day.Moreover,social networks have completely reshaped people’s life and work styles.Business recommendation,public opinion control,rumor blocking and other social needs make the information diffusion on social networks become a hot topic at present.The difficulty of studying such problems lies in the complexity of human behavior,which is not only difficult to quantify the factors affecting information diffusion,but also unable to determine the mode of information diffusion.Many researchers,according to their own ideas,have established the information diffusion model.Most classic social network diffusion models need a complete network structure,in which the connection attribute between users is also available.Generally,however,in real social networks,the number of users is so huge that it is difficult to accurately describe its specific network structure.By looking back on the real life,when we study the interpersonal relationship network between people,the network topology we obtain is usually incomplete and fragmentary,which makes it very difficult to predict the diffusion of information.Therefore,in order to overcome the problem of incomplete data,the method of representation learning is introduced.On the basis of defining the factors that affect the information diffusion,it generalizes them with mathematical models,and finally restores the dynamic process of information diffusion as much as possible.To explain the mechanism of information diffusion on social networks,we introduce the concepts of complex network theory and information propagation theory.We took four famous social networks as the research object to analyze the relationship between the characteristics of online social networks and information diffusion.Moreover,based on the basic theory of complex network,the relationship between network topology and information diffusion behavior,such as user publishing and forwarding,is studied from the perspective of degree distribution,assortativity,small world and scale-free characteristics.We attempted to find out the potential factors that affect propagation on social networks.We then introduce the idea of representation learning and related models,and proposes a social network information propagation model(IPM)based on user’s preferences and influence.The model synthetically considers the two factors of user influence and user interest,and automatically embedded the user and propagation item as a low dimensional vector in Euclidean space,and then predicts the propagation result according to the value of eigenvector.Every user in the network is projected into a latent space,which is called influence space.The degree of influence between users is derived from the geometric distance between their eigenvectors,that is,the closer the distance is,the stronger the influence is.At the same time,all users and propagation items are projected into the second latent space,which is called user preference space.In addition,the user’s interest in the propagation item is derived from the geometric distance between their eigenvectors,that is,the closer the distance,the greater the degree of preference.In the algorithm,the expectation maximization algorithm is used as the basic framework,the random gradient algorithm is used to optimize the value of the eigenvector,and the parameters of the model are settled through simulation experiments.Finally,we designed two schemes and three indexes to evaluate the model.The prediction experiments of information transmission are carried out in multiple artificial networks and real networks.We found that the accuracy and time cost of the model are also better than other models without network structure,which shows that it can more accurately simulate the propagation process and predict the results of information diffusion.All in all,after determining the factors that affect information diffusion,we use the method of representation learning to avoid tedious feature engineering,enrich the physical meaning of the research object,combine the relevant external features,improve the accuracy and efficiency of information diffusion model in prediction results,and provide a novel way to explain the information diffusion mechanism.