Research On Point-of-Interest Recommendation Algorithm Based On User Dynamic Preference And Attention Mechanism

With the popularization of intelligent devices and the rapid development of mobile Internet technologies,location-based social networks(LBSNs)are becoming common and prevalent.As the core intelligent application of LBSNs,Point-of-Interest(POI)recommendation can predict the next POI that a user will visit based on his/her historical mobility trajectory recorded on social networks.However,existing studies on this technology still faces two challenges.Firstly,a user's historical check-in trajectory contains sequential transaction patterns of the user in the physical world,which is crucial for predicting the user's next check-ins.However,the sequential patterns of users are often high-order and have complex spatiotemporal dependencies,which are hard to be explored effectively.Secondly,users' preferences in the long-term check-in history are complicated and dynamic,and it is very important to concurrently model users' long-and short-term preferences.However,existing methods are difficult to effectively learn users' dynamic long-and short-term preferences and ignore the mutual effect between these two preferences.To address the above challenges,the work of this paper mainly includes the following three aspects.(1)To effectively capture complicated sequential transaction patterns in users' check-in behaviors,a spatiotemporal self-attention network based sequential preference model is proposed.The model utilizes the time interval and geographic distance information among POIs of the check-in sequence to construct relative time embedding and geographic attention weight respectively to improve the standard self-attention network.The network can fully learn the complicated and fine-grained transfer dependencies(i.e.,semantic,temporal,and spatial relationships)between any pair of POIs within the check-in sequence.(2)To further learn the complicated and dynamic preferences of users,a memory augmented hierarchical attention network is proposed based on the spatiotemporal self-attention network,which can dynamically model both long-and short-term preferences of users.For long-term preference learning,we employ key-value memory network to maintain fine-grained latent preferences of users.We design the read/write operation of memory network,which enables the network to dynamically manage users' long-term memories according to their constantly updated check-in records.For short-term preference learning,Finally,we design a novel long-short co-attention network to capture the potential correlation and mutual effect between long-and short-term preferences.(3)We conduct extensive experiments on two real-world LBSN datasets and compare the proposed method with the existing representative recommendation methods.The results show that the proposed method has significantly improved the performance of multiple evaluation criteria,which proves the effectiveness of these proposed models.At the same time,each component of the proposed models is validated by ablation experiments.