Mining Social Influence In Social Media

In recent years,social inf luence analysis has become one of the important research topics in social media.Social media enables Internet users to communicate with others anytime and anywhere.Analyzing social influence in social media can help promote the application of social influence in information dissemination,recommendation system,link prediction and expert discovery,etc.However,with the popularity of social media,the requirements for various applications are increasing,the same users often appearing in multiple social media,and mining group influence becoming an emerging issue.The existing research works cannot meet the needs of these applications.In this paper,we further explore a number of issues in this important topic,focusing on how to measure the distance between users in multiple social media and how to model and represent the social influence of individuals and groups in social media.Specifically,the main contributions of the dissertation are summarized as follows:1.Jointing Metric Learning and Boosting for User Distance MeasurementIn order to model and measure user distance in multiple social media,we try to address the data sparsity problem in single social media by taking into account individual attributes and network structure of users in different social media,and taking into account user characteristics,link information and network topology.We construct a metric learning model and propose an adaptive metric learning algorithm based on metric learning and Boosting framework.The algorithm learns relevant knowledge from relevant social media to help distance metric learning in the embedded feature space of social media,and employs Boosting technique to eliminate irrelevant attributes.In particular,our model can avoid the over-fitting problem.The link prediction experiments are carried on two real large-scale datasets.The results verify the feasibility and effectiveness of the proposed model and algorithm.2.Mining Individual User Influence Based on Electromagnetic Field TheoryThere are a variety of definitions of individual inf luence and methods of calculating inf luence for different experimental purposes,but they can not effectively process and accurately evaluate individual inf luence in Weibo.In this paper,based on the concept of source in electromagnetic field theory,we propose new concepts such as positive microblogging source,negative microblogging source and neutral microblogging source in the field of microblogging.Then,we propose a new Individual Influence Rank Algor ithm(IIRank).The algorithm first uses the method of calculating magnetic flux to calculate the microblogging flux according to the behavior information of the microblog user,to evaluate the user's behavior.Then,the microblogging flux is used to calculate the microbloging flux density to rank the user influence.On the real microblogging dataset,we experimentally verify the validity of the proposed model and algorithm.3.User Interaction Analysis based on Game TheoryAiming at the interaction of users,we propose a user interaction prediction model based on the game theory,which uses a hybrid strategy Nash equilibr ium to predict the user's attitude by using the postings and replies.In order to further study the interaction of users in the real social media,analyzing the forwarding behavior of users is the key to the understanding of information dissemination.Furthermore,we propose a user interaction prediction model based on hybrid strategy game to analyze the interaction mode of users in social media.The model can predict the user's forwarding behavior.We verif ied the validity of the model on a real weibo dataset.Finally,our model provides a feasible approach for constructing kernel functions.We propose a game kernel function SVM classification algorithm.The algorithm can effectively integrate the advantages of local kernels and global kernels to obtain better classification results.Experiments on some standard datasets verify the effectiveness of the proposed model and algorithm.4.Mining Group Influence Based on Multi-attributionIn response to the emerging issue of modeling group inf luence,we define some concepts about the group influence,formalize the modeling of group influence,and build a group inf luence analysis model that combines user influence,social trust,and user engagement in the community.The model first removes the zombie fans and then calculates the user's influence.Then,the user's final inf luence is calculated by combining the user's individual influence and the willingness to diffusing thematic information.Finally,by using the final inf luence of the users in the community,social trust,and the closeness of users to assess group influence.On the real Sina Weibo data,we have designed a Community-level Influence Analysis Algorithm(CIAA).Experiments verify the effectiveness of our proposed algorithm.