Research And Implementation Of A Probabilistic Inference Model For User Behavior Relationship On Social Network

Social network is an application which focuses on users and generates a variety of information from users. The information is the basis to observe the relationship between the user. The method that finding users who have close relationships and providing information or recommendation service for them, and then serving the community, is one of the hot issues about the study on the relationship between users on social network in academia and industrial now. As one of the typical social network applications, with many users and containing rich content of Tweets, Twitter can be effectively mined user relationships. In this paper, we have a deep study of user relationships from user behavior’s respective based on Twitter,.First of all, most of researches about user relationships only focus on the study of "follow" and "followed",but the study can not definite reflect close relationships between users. In this paper,user relationships can be analyzed by two key factors: activity and location and then Behavior Relationship is presented.Behavior Relationship(BR) is that users conducting similar activities in close locations,such as meeting, traveling, shopping and so on. Further, BR is divided into User Pair Behaviors Relationship(UPBR)and User Group Behaviors Relationship (UGBR).Secondly, "@" on Twitter reflects close degree between user relationships, so a Probability Inference Model(PIM) is proposed through "@".The model firstly computes probability of activity similarities and location similarity for a pair of users,and then inferences whether they have UPBR by probability method. Secondly, whether the user groups will have UGBR can be found out by BR Matrix-based Maximal Tree Clustering(BRMC).Finally, a scalability experiment was carried out on the effects of BR inference from two data sets which are real data and simulation data. Among them, real data test the error distance, accuracy and the relation between @ times and error distance on inferring locations.The experimental results show that PIM has a relatively high accuracy on inferring locations. There are two aspects for simulation data, one is testing accuracy, recall and F1-Measure of PIM on UPBR inference under different activity threshold;Another is testing the clustering accuracy of BRMC based on NMI, F1-Measure index and testing BRMC related parameters from the clustering sensitivity perspective,the experimental results show that BRMC has a fine effect.on UGBR inference.