User Behavior Analysis And Location Prediction In Location-based On Social Networks

With the popularity of GPS(Global Positioning System,GPS)devices and location-based mobile social networks(Location-based social networks,LBSNs),massive trajectory data is constantly accumulating in our daily lives.Because this type of data contains features such as temporal feature and spatial feature and semantics feature,such data is critical for analyzing user behavior patterns and predicting the user's next location.Based on the research on related literatures in recent years,although there are many studies to predict the user location by mining the spatial and temporal features of users,if considering the original trajectory data lacks the semantics of the location,this paper first proposes an algorithm which combine of semantic enhancement and spatial-temporal features to predict user location.Then,an urban "hot zone" mining algorithm is improved.Finally,a prototype system for mining big data of LBSNs is constructed.Firstly,by studying the multi-dimensional features in LBSNs,a stay points semantic algorithm and urban "hot zone" mining algorithm are proposed.Among them,the semantics of the stay point improves the accuracy of the user's position prediction and accelerates the convergence speed of the model training.The "hot zone" mining algorithm is verified according to the contour coefficient,and the clustering effect is better than the benchmark model.Under the same data set,compared with the results of similar algorithms in the past research,the urban "hot zone" mining algorithm in this paper mines a more fine-grained "hot zone",which improves the range accuracy of the position prediction in this paper.Secondly,this paper analyzes the user behavior in LBSNs and constructs an SSTAN(An Semantic reinforcement and Spatial-Temporal Attention Networks)to integrate the multi-dimensional features.This paper proposes a geographic location prediction algorithm named GLP-SSTAN(Geographical Location Prediction algorithm based on Semantic reinforcement and Spatial-Temporal Attention Network)to make location prediction in LBSNs.Then,based on the staying point extraction algorithm,urban "hot zone" mining algorithm,SSTAN and GLP-SSTAN algorithm,this paper constructs a prototype system named NUPT ST-Data Miner for mining user behavior and location prediction in LBSNs.The main functions of the prototype system include visualization of the stay points,visualization of the "hot zone",visualization of the user's location prediction,and recommendation services around the stay points.Compared with the benchmark model in the related literatures,the prototype system has three advantages: good visual visualization,kindly user interaction and addable functional modules.Finally,experiments in this paper were conducted based on the real data set named Geo-Life provided by Microsoft Research Asia.The results show that the urban "hot zone" mining algorithm has better clustering effect than Mean-Shift and K-means algorithms(the contour coefficient is better);GLP-SSTAN algorithm is superior to the benchmark model in terms of accuracy,recall rate,location prediction accuracy,and model applicability.