A Spatio-temporal-preference Graph Attention Network For POI Recommendation

The rapid development of Internet has given rise to a series of mobile location-based services platforms.Based on massive user behavior data,these platforms use data mining methods to provide users with intelligent life services and advice.Among these services,personalized POI recommender system plays an important role,providing great convenience for people's daily life.POI recommender system algorithm is a research hotspot and has attracted the attention of a large number of researchers.Based on the user's historical check-in data,existing research works design matrix factorization,deep neural network,convolutional neural network and other methods to model the user's interest,and then make personalized recommendation.However,most of these models have defects such as incomplete user interest modeling and insignificant spatiotemporal characteristics.As a result,they cannot accurately recommend items to users.Therefore,after thinking about the POI recommendation scenario,this paper puts forward three modeling directions for users' interest: personal behavior preference,temporal factor and spatio factor.Based on the three directions,this paper carries out the research on the POI recommendation model.In terms of model design,this paper focuses on graph neural network,which has developed rapidly in recent years.Based on the idea of divide and rule,this paper carries out in-depth modeling of three types of user interests from the perspective of multi-source graph.Specifically,three types of multi-source graphs are constructed based on the user's checkin behavior data: preference tranfer graph,temporal correlation graph and spatio correlation graph.Then,based on the multi-source graphs,a spatio-temporal-preference graph attention network(STPGAT)for POI recommendation is proposed.Finally,an end-to-end training paradigm and prediction flow are carried out to realize personalized POI recommendation.Specifically,STPGAT mainly uses multi-layer graph neural network to explore the higherorder connectivity between users and POI nodes in the information transmission in the graph,and uses the attention mechanism to focus on the targeted aggregation process of surrounding nodes to central nodes.At the same time,STPGAT introduces the information fusion method based on multi-task learning,which effectively realizes the organic integration of user behavior preference,temporal and spatio factors.In terms of experimental verification,this paper selected three real world POI recommendation data sets from small to large scale(each data set contained a large number of historical check-in behavior records of different users),and carried out a detailed comparison experiment between STPGAT,the classic recommendation model and the starte-of-the-art(SOTA)model.It is proved that STPGAT has the best performance from various indexes.In addition,each component of STPGAT has been ablated to verify the effectiveness of the design of each component.Finally,based on the behavior records of a specific user,a quantification experiment of POI recommendation interpretability based on multi-source graph structure is designed tentatively,which proves the practicability of the proposed STPGAT model in the field of POI recommendation from three aspects of user behavior,time and space.