Study And Application Of User Linkage Across Social Network And Spatio-temporal Database

The development of Global Positioning System,sensor techniques,and mobile devices has generated a large number of trajectories of moving objects.In addition,the flourish of social networks,such as Sinaweibo,Wechat,Foursquare,Twitter,and Instagram,has provided many platforms for users to share ideas and communicate with each other,and has generated tremendous social network data.Obviously,we can obtain more abundant data for a user after linking the same accounts belonging to him across spatio-temporal database and social network.Then,we can construct a more comprehensive profile for the user to support many real applications,such as location prediction,time prediction,topic prediction,to name a few.Specifically,we investigate three problems in this paper.(1)Linking cross-platform users based on density based methods.To link users across different platforms,we need to extract the features of corresponding users before measuring the similarity between two users.Compared with computing the user similarity with same records,the clustering based method is more effective.In spatial domain,we use the density based clustering method DP to find the distribution of stay regions.In temporal domain,we use the Gaussian Mixture Model to find the global time distribution and local time dis-tribution of users.To further improve the performance of the proposed method,we assign different weights to all extracted features based on the idea of TF-IDF,where the discrim-inate features are assigned with large weights,and the common features are assigned with small weights.The real-world datasets based experiments demonstrate that the proposed method outperforms the existing studies.(2)Linking cross-platform users with sparse datasets.The clustering based method is effective on dense datasets,yet the effectiveness of the method declines fast on sparse datasets.Unfortunately,social network datasets are usually spare and unbalanced,as many users do not provide check-in information while sharing status,due to privacy consideration.In this case,it is hard to extract the distribution of stay regions.To tackle the problem,we proposed a kernel density estimation(KDE)based method,where the similarity between two users are directly measured with historical records.However,the KDE is inherent inefficient due to the high computation complexity.To improve the efficiency of the method,we use grid map to organize spatial information and divide the time into different bins.Then,users are presented by a set of grid cells and bins with corresponding probability.In addition,to compute the similarity between two users more precisely,we assign different weights to grid cells and time bins based on renyi entropy.The experimental results demonstrate that our proposed method that considers grid cell,time bin,and entropy simultaneously can achieve both high effectiveness and efficiency.(3)Constructing user profiles and investigating the applications of them.Following the cross-platform user linkage,we can obtain more abundant information and construct a comprehensive profile for a user.Note that,the profile contains the stay region distribution,transfer behaviors between these regions,time distribution,and topic distribution in each region.To investigate the application of the profile,we present the performance of our method while predicting locations,time,and topics.The experimental results demonstrate that the constructed profile has high performance in prediction.Finally,we summarize the study and present the future work.