Research On Influence Analysis In Social Networks

The in-depth development of information technology,coupled with the widespread use of smart terminals,make social networks an important medium for people to publish and obtain information.With social networking platforms,various complex relationships among people in reality can be embodied and extended within the virtual network.Meanwhile,the information about events occurring in the real world will spread along with the online interactions between users,which in turn affects their offline behaviors.Users are both the producers and disseminators of information,and their influence is not only the major driving force of information dissemination and structure evolution,but also the main inducement of social behavior formation.In fact,opinion leaders often leverage their influence to attract public attention quickly and guide the dissemination and development of public opinion events by posting information containing their own personal opinions.In the process of information dissemination,the influence strength of social relationships determines the propagation path of information,which indirectly affects the rate and scope of public opinion events in the cyberspace.With the interactions between users,such as reposting,social influence propagates through information,and the effect of public opinion will affect the development of the event.Therefore,there is of significant practical value in studying social influence in order to monitor public opinion and safeguard national security and stability.Due to the wide participation of users,the complexity of social relationships and the rapid dissemination of information,social networks have unique characteristics,such as massive data,diverse topics,complex interactions and structural evolution,which bring both opportunities and challenges to social influence analysis.Based on existing researches,this dissertation mainly studies the causes of influence,the influence strength of users and the effect evaluation of influence diffusion in the process of information dissemination.The main contributions are as follows:(1)In terms of influence model construction,we proposed a unified framework for measuring user influence based on the random walk model.Existing influence algorithms mainly focus on describing user influence from a single perspective,thus failing to reflect complex factors and mechanisms comprehensively and evaluate the impact of different factors with social networking data.We incorporate three important elements into the unified framework,including user,information and network structure,and the generalization ability of the framework can be significantly enhanced by parameterizing different factors.With big data processing frameworks,we conduct extensive experiments on a real large-scale dataset to test the effects of relevant parameters on the performance of the classical instantiated influence algorithm.(2)In terms of individual influence measurement,we proposed a topic-level influence measurement based on user behavior.Existing researches tend to calculate a global influence for a user,and rarely distinguish their influence across different topics.Since social behavior is the external form of influence,it can be used to calculate user influence directly.By analyzing the temporal patterns of user post and repost behaviors,we can capture the behavior related influence propagation coefficient.And the topical influence propagation coefficient can be achieved with information content via LDA model.Then with the linear weighting method,the proposed method can be used to calculate topic-level influence for each user based on the unified framework above.Experimental results on Sina Weibo show that our method can effectively mine influential users in different topics and be convenient for parallel computing.(3)In terms of mutual influence analysis,we proposed a novel mutual influence model for information dissemination.Most previous works discuss mutual influence by analyzing the influence strength of social relationship between each pair of users,which is prone to causing overfitting problem and reducing the accuracy and efficiency.Considering user is the key factor of information dissemination,we associate each user with an influence vector and a susceptibility vector.And the mutual influence between users in different topics is the product of the corresponding elements of the two vectors above.The proposed model can be directly used to determine the propagation probability in regard to specific information.Experiments on synthetic and real datasets demonstrate our model has better performance than other methods in mutual influence measurement and propagation probability calculation.(4)In terms of repost behavior prediction,we proposed a mutual influence based repost behavior prediction method.Traditional methods mainly use social characteristics,text attributes and historical information to predict user behavior,without considering the time factor and social influence.To this end,a novel algorithm for calculating timesensitive mutual influence is developed,and together with another two features of interest coverage and information quality,we train a classifier for predicting repost behavior with the logical regression model.The experimental results on Sina Weibo show that compared to other methods,our method can achieve better prediction performance while guaranteeing computational efficiency.In conclusion,this dissertation mainly focuses on influence analysis from users' perspective in social networks and provides some technical solutions.Extensive experiments on both synthetic and real datasets validate the effectiveness and efficiency of our proposed methods.Our work is of great theory and application value on social influence analysis.