Research On Point Of Interest Recommendation Algorithm Based On Attention Mechanism And Graph Neural Networ

With the rapid development of Internet technology and the popularity of mobile devices,Location-Based Social Networks(LBSNs)are gradually known and used by more and more users.Hundreds of millions of users check in to share their points of interest(POI)on LBSNs.A large number of check-in data not only brings convenience to users’ travel choice,but also causes a very serious problem of information overload.Users is very difficult to find which POI is they really interest in from numerous POI.As a significant method to effectively solve the information overload problem,POI recommendation is widely applied in the LBSNs platform.Although POI recommendations have been widely studied and applied by academics and industry,there are still some problems with the existing methods: first,existing methods only consider the spatiotemporal correlation between adjacent check-ins,ignoring the spatiotemporal correlation between non-adjacent check-ins;second,existing methods learn users’ preferences only from their own historical check-in records,which leads to bad recommendation performance in sparse datasets;third,existing methods use supervised learning models for recommendation tasks,but the sparsity of supervised signals and noise make them difficult to learn high-quality representations of users and POIs;fourth,existing methods are unable to integrate information from different domains effectively.In order to solve these problems,this paper leverage attention mechanism and graph neural network to improve the performance of point-of-interest recommendation,and proposes the following three algorithms:(1)POI recommendation algorithm based on spatio-temporal interval attention mechanism named STIA.Firstly,the spatial-temporal interval attention encoder is used to learn the spatial-temporal correlation between check-ins in the user trajectory and update the representation of the user trajectory.Then,the spatial-temporal interval attention network is used to aggregate the user trajectory according to the spatial-temporal correlation between the candidate POI and check-ins in the user trajectory,so as to learn the user’s preferences.The experimental results show that the proposed method can take full advantage of the spatial-temporal information in the data and obtain a more ideal recommendation effect.(2)POI recommendation algorithm based on heterogeneous graph neural network named LSPHG.In the long-term preference learning model,a heterogeneous graph is constructed to represent the check-in information of all users and the category information of POIs.Then,the heterogeneous graph neural network is used to capture the higher-order relationships among nodes in the heterogeneous graph and learn the representation of users’ long-term preferences.In the short-term preference learning model,the STIA model is used to learn users’ short-term preferences.Finally,the user’s long-and shortterm preferences are linearly combined with personalized weights to predict which POI he will visit.The experimental result show that the proposed method can better learn users’ preferences and effectively improve recommendation performance.(3)POI recommendation algorithm based on graph constructive learning named GCLSGR.In the recommendation task,the expressions of users and POIs are learned through the recommendation encoder,and the list of POIs is recommended to the user.In the contrastive learning task,domain-specific contrastive learning is used to enhance the representation of entities in each domain,and cross-domain contrastive learning is used to transfer knowledge from the social domain and geographic domain to the check-in domain.Finally,the joint training strategy is used to optimize the model.The experimental results show that the proposed method can more effectively utilize users’ social relations and geographical location of POIs to solve the widespread problems of data sparsity and cold start in POI recommendation and obtain good recommendation performance.