User Behavior Modeling And Research In Location-based Social Network

Recently,with the fast expansion of mobile Internet and maturity of position tech-nology,Service platform and information which is related with Location-based Social Network(LBSN)has been widely used in our life.A large amount of Location-based datasets is accumulated by the widely use of Location-based Service(LBS).This pro-vides strong and firm support for mining the hidden user preference.It will be more benefit for their living and traveling via analyzing rule of user's behavior and prefer-ence.Also,the analyzed results about user preference could show merchant and rel-evant deciders with meaningful suggestions and instructions.Thus the focus of this paper is to mining user's preference from the current and prospective view to give rec-ommendation and prediction in accordance with users' interest and current states.Cor-responding problems can be boiled down to Point-of Interest(POI)Recommendation and Location Prediction.Although LBSN provides sufficient data resources,it is chal-lenge for these two problem by using present methods since heterogeneity and sparsity in LBSN datasets.Facing with characteristics and challenges existed in LBSN datasets,we propose corresponding methods to solve the problem in POI Recommendation and Location Prediction.Specifically,they can be seen in follows:1.As for POI recommendation,this paper proposes a mixed POI recommendation model based on heterogeneous information.Abundant entities and relations among entities are contained in LBSN,which are regarded as heterogeneous information in Location-based datasets.It will be better to increase the effectiveness of POI rec-ommendation via integrating these information effectively using reasonable modeling and algorithm design.Aiming at the heterogeneous information such as user's rating and textual information,social correlation and geographical information,we propose a mixed POI recommendation method based on the combination of virtual interest of users and realistic distance between users and POIs to combine above three information organically.Specifically,we use kernel density estimation to measure the geographical distance among locations,collaborative filtering method which is based on friend and users with common check-ins to exhibit users' psychological perception.Also we use users' and POIs' reviews aggregated Latent Dirichlet Allocation(LDA)to mining the preference of users and POIs and model the explicable aspect of users' virtual interest.Correspondingly,we use probabilistic latent factor model to modify the inexplicable as-pects of users' virtual interest.Finally,we combine above modules into a mixed model for POI recommendation organically.Experiments indicate the superiority of mixed model of POI recommendation comparing with state-of-the-art POI recommendation methods.Also,this model shows superiority on prediction and robust.2.As for location prediction,this paper proposes a location prediction model based on check-in sequence via latent topic embedding model.Researches show that user's behavior in LBSN is regular and predictable.What's more,it is also related with the circumstances of users and locations.For most users,their check-ins are relatively high sparse comparing with the whole distribution of datasets.So how to make location pre-diction with the overall consideration of above characteristics is an urgent problem.This paper proposes a latent topic embedding method based on check-in sequence.Specifi-cally,We use region-based gaussian distribution model to model the geographical infor-mation in LBSN.To remit the sparsity of social relation,we expand user's relations.At the same time we combine context-aware word embedding and temporal LDA to model the situation of users' check-in behaviors.As for the regularity of their check-ins,we split continuous time as time pattern via horizontal and vertical way which discretizes the continuous time.We can fetch users' preference and locations' embedding by uti-lizing above methods so that we can make location prediction on the next time pattern efficiently.Experiments on a typical dataset show more accuracy comparing with tra-ditional methods in location prediction.