User Behavior Prediction And Point Of Interest Recommendation Based On LBSN Check-in Data

With rapid development of location-based social network, massive location data accumulated enables us to provide effective personalized POI(Point of Interest) recommendation. Essencely, human movement behavior is spatiotemporal sequence from one <location, time> transfer to another <location, time>. However, existing research did not fully exploite the characteristics of spatiotemporal sequence. To extract its characteristics in-depth, this paper proposes the AMGR(Additive Markov chain and Gravity model based Recommendation system), which combines Additive Markov Chain for prediction of user spatiotemporal sequence with Gravity model for POI recommendation. The three main works of this paper are as follows:(1) User behavior prediction based on Additive Markov Chain. First, spatiotemporal sequence extracted from the historical check-in data is modeled as LLTG(Location-Location Transition Graph). LLTG is a real-time online model with constant time complexity of incrementally updating the check-in data stream. Then, Additive Markov Chain is applied to predict the probability of spatiotemporal sequence probability with higher accuracy than first-order Markov Chain and with higher efficiency than traditional n-order Markov Chain in prediction problem.(2) POI recommendation based on Gravity model. This paper introduces the Gravity model commonly used in the field of city traffic, which considering the spatial, temporal, friend relationship and POI popularity to calculate the attractive force between POIs. The attractive force from gravity model is used as the weight of Additive Markov Chain to get the user's final visit probability of new POI, and recommends the top-k POIs with the highest probabilities for the user. With the combination of the Gravity model and Additive Markov Chain, the AMGR system achieves the recommendation model that comprehensively integrates spatial and temporal factors, friend relationship, POI popularity and sequence factors.(3) Evaluation of AMGR. Experiments are conducted on large volumes of real check-in data collected from Gowalla and Brightkite to evaluate the recommend accuracy and recall of AMGR system. Experimental results show that the AMGR's accuracy reaches 0.31 and 0.22 and the recall reaches 0.21 and 0.12. Compared with latest location recommendation methods, the AMGR achieves meaningful recommendation results due to significantly increased accuracy and recall.