Research On Friendship Prediction Based On Location-Based Social Network

With the prevalence of social networks and the development of mobile smart devices, location-based social networks(LBSN) has gradually integrated into people’s lives. LBSN aims at combing the users’ mobile behavior and geographic location information, to fully dig out the correlation between potential behavior and their daily activities geographically, establishing a close link from virtual network to real-life.Thus it undoubtedly provides new research directions in social relationship prediction and other fields. LBSN records check-in information in time and spatial dimensions, making it possible for friendship prediction.However, it’s noted that the network structure in LBSN is relatively sparse, difficult to completely reflect individual movement behavior.Accordingly, it has become an important topic that digging out the user information, and utilizing the implicit knowledge to portray users’ characteristics. To realize the above objective, the paper proposes a friendship prediction framework based on Support Vector Machine(SVM), and exploits Gowalla and Brightkite datasets to simulate experiments. Some results show our algorithm valid to some extent. Morespecifically, the main work of this paper includes the following four aspects:Firstly,exploiting check-in information to analyze user’s behavior characteristics in Gowalla and Brightkite datasets, such as the number of the user’s friends, check-ins and check-in location. And we found that it follows a long tail distribution and the number of friends over 50 accounts for a very small proportion. The check-ins is less than 10 times in Brightkite reaches 43.5%, further illustrating the sparsity of data;posing new challenges for friendship prediction.Secondly, in order to dig out the principle of user movement, we analyze the area and periodicity of user movement by virtue of check-in information in time and space. From the study of moving area, we can see that most users only move within a relatively small range. Generally speaking, the active radius of users in Gowalla dataset is higher than in Brightkite. Moreover, the periodicity of users’ mobile trajectory can reflect the regularity of life to some extent.Additionaly, extracting the relevant features on the basis of the analysis of network structure and user behavior. According to the conventional methods of node similarity, this paper proposes an approach to illustrate the user social relationships, and extracts the user check-ins and check-in type as key features for prediction.Finally, establishing the prediction model based on SVM, integratingthe above features for friendship prediction, and evaluated by accuracy,recall, F1-measure and AUC value. It can be seen that social relationship makes a greater impact on the friendship prediction than others. The accuracy based on three key features is higher than that only on one feature. To further improve the prediction results, the Genetic Algorithm(GA), Particle Swarm Optimization(PSO) and Grid Search(GS) on can be used to optimize the penalty factor C and kernel parameterg.