Research On Influence Maximiation In Social Medias Based On Information Diffusion Network Embedding

Online social media has become the main channel for people to exchange information and communicate due to its immediacy and interactivity.When users post messages on the social media,their followers can receive and forward the messages.Therefore,users can influence others and make them generate some thoughts or take some actions through such information diffusion process.Research on the law of information diffusion helps to predict the behavior of individuals or groups,and provides guidance and suggestions for professionals in marketing,content recommendation,public opinion control,and etc.The influence maximization problem is aimed at finding a certain number of nodes in the network under a given influence propagation model to maximize the influence spread,i.e.,the expected number of nodes who are influence by these selected nodes.With the rapid growth of the number of users and the vigorous development of social media platforms,the large-scale and complex structure of social network data has increased the difficulty of analyzing and modeling the influence maximization problem in social media.This dissertation proposes a research method of influence maximization based on the network embedding of the information diffusion network.The method uses the efficient representation of network embedding on large-scale and complex graph data to capture the unique characteristics of influence diffusion among users.The main research work and contributions of this dissertation are as follows:1.This dissertation made a long-term observation of the social media followership network,analyzed the evolution law of its main characteristics,and explored the mutual influence between the information diffusion and the evolution of the network structure.This research provides support for constructing a co-evolution model of information diffusion and network structure,and then more accurately simulates the process of information diffusion in social media,and finally identifies influential users to maximize the influence spread.2.Aiming at the problem of target user mining in the task of influence maximization in social media,this dissertation proposes a specific user classification method based on heterogeneous information network embedding to identify high-value users corresponding to the target task.This dissertation first builds a heterogeneous information network model of social media data.The relationship between various types of users is expressed by the meta-paths.After that,the classification methods based on transductive learning is applied to measure the value of different meta-paths for user classification,and pick out the high-value meta-paths.Through network embedding based on a meta-path,the vector representation of each user is learned.The attention mechanism is used to automatically learn the weight of each vector generated from different meta-paths,such these vectors could be aggregated to the final user feature vector.Lastly,users are classified based on the similarity of the feature vector.This method is proved to be more accurate than state-of-the-art user classification methods.3.Aiming at the difficulty in designing the parameters of influence diffusion model in social media,a method for estimating influence strength based on the embedding of information diffusion network is proposed.This dissertation considers that users can play two different roles as sender and receiver in the process of information diffusion in social media.Through network embedding,the different characteristics of each user when playing as sender and receiver are projected into two different vectors with identical dimensions.On this basis,the asymmetric influence strength between users is calculated,and the influence propagation model with the corresponding parameters are designed.This method can quickly estimate the probability of users being affected by multiple seed users,and improve the efficiency of identifying influential users.4.Aiming at the problem that the characteristics of information diffusion across multiple social media paltforms are complicated,a method for constructing an influence diffusion model based on multiplex network embedding is proposed.There are different patterns of the intra-network propagation and inter-network propagation of information,so a heterogeneous diffusion model is constructed.This dissertation proposes a random walk strategy in multiplex networks,and applies a random walk-based multiplex network embedding method to project the user's characteristics in each network into different vectors,and designs parameters for the heterogenous influence propagation model accordingly.On this basis,a group of users with high-value influence in multiple social medias is identified.The experiment shows that influential users in multiple networks are quite different from influential users in a single network.