Information Fusion And Diffusion Pattern In Multiple Social Networks

People use various social platforms with different features to meet their different needs and purposes.In this study,users who have personal accounts in multiple online social networks are called “overlapping users”.Compared with the single social network,expanding information in multiple social networks in breadth and depth leaves big challenges for researchers to study the multi-dimensional information fusion and diffusion over multiple social networks.Based on the concept of information fusion,this thesis analyzes and studies the elements of“user”,“network” and “message” in social networks for multiple social scenarios to form three problems.The first research is the topic preference mining of users in multiple social networks which help us understand what the user’s preferences exactly are.Community detection across social networks is the second research,which can figure out users with whom users are likely to interact.The third research is the information diffusion across social networks to predict which messages users will forward.In particular,how to take advantage of the data of multiple social networks to profile users and how to dig out the differences and connections of the characteristics of users in different social networks are key points in the first research.The problems to be solved in the second part include how to define a reasonable concept of community across social networks and further construct such communities with intrinsic user connections.The challenges in the third research are how to model the phenomenon of information diffusion between social networks,the route of forwarding messages,and the potential factors affecting the diffusion of information.To this end,the research content and contributions of this thesis can be summarized as follows:(1)This research proposes a topic-preference discovery method for multiple social networks,which can fuse data from various social networks for unsupervised discovery of user topics.First of all,user topics are divided into two types: global topics and local topics,based on data from multiple social networks and a single social network respectively.Then,user behavior data across networks are incorporated to build the user topic-preference model.After the calculation of model parameters through the Gibbs sampling algorithm,the topicpreference features of users are obtained,and further used to analyze the characteristics of social networks and the intrinsic motivation of users to use these networks.Finally,the comparison experiments with the existing works are conducted on the dataset of three real social networks to verify the effectiveness of the proposed method.The evaluation metrics of the experiments are rich,such as perplexity,likelihood,and word mutual information score.(2)This thesis presents a global community detection mechanism across social networks based on overlapping users.It leverages the alignment matrix to distinguish overlapping and nonoverlapping users of different social networks.Then,taking the overlapping users as the core,the interaction attribute matrix between users can be constructed based on the topological structure and content similarity,and the communities across social networks can be discovered by the multi-constraint non-negative matrix factorization method.This model can not only detect the community from the perspective of every single social network but also obtain the global community that integrates multiple social networks.Finally,the experiment on the dataset of three real social networks is conducted to verify the effectiveness of the model in terms of community quality and community integration,based on the evaluation metrics for communities proposed in this thesis across multiple heterogeneous social networks.(3)A cascade model of information diffusion for multiple social network scenarios is proposed in the thesis.First of all,taking user preferences,social relations,platform environment,and external influences into consideration,it models the factors affecting the production process of information and constructs the posting generation model of users.Then,considering that the user’s choosing sequence of social networks is related to the information diffusion,a platform choice model of users is also built in this thesis.The posting generation model of users and platform choice model of users jointly describe the posting process of users across multiple social networks.After that,the model parameters are calculated by the Poisson process-based cascade model of information diffusion and the Classification-based Expectation-Maximization algorithm.With the potential semantic correlation between users’ posts mined by the above model,the cascade structure of information diffusion is further modeled based on time series.Finally,the cascade structure of information diffusion output by the model is analyzed,and the prediction ability of the model for information diffusion is estimated through experiments.