Research On Point-of-interest Recommendation Models Based On Deep Learning

In recent years,Point-of-Interest recommendation has become a hot research field in both academia and industry with the gradual rise of Location-based Social Networks.There are various deep learning-based models proposed for POI recommendation that have achieved significant results.However,existing deep learning-based POI recommendation models still have some problems.First,existing models only focus on the local spatio-temporal relationship of the current user trajectory when modeling the relationship between POIs,without comprehensively considering the global spatio-temporal relationship of all user trajectories.Second,existing self-attention network models do not sufficiently explore the personalized spatio-temporal relationship of user trajectories and lack a personalized representation method for user spatio-temporal information.Finally,the existing graph neural network based POI recommendation models mainly construct homogeneous graphs of user-user and POI-POI,or simple heterogeneous graphs of user-POI,which have single graph structure and limited information expression ability.This thesis conducts in-depth research on the above-mentioned problems and proposes corresponding models to solve the shortcomings in existing research.The main contributions of this thesis are as follows:1.A global spatio-temporal aware graph neural network model is proposed.In order to comprehensively consider the global spatio-temporal relationship of all user trajectories,this model firstly breaks the independence of all user trajectory boundaries and constructs a global spatio-temporal graph.Then graph neural network is used to model the global embedding vectors of POIs to model the global spatio-temporal relationship between POIs.2.A spatial-temporal weight matrix is proposed.In order to fully explore the personalized space-time relationship of user trajectory,the proposed matrix converts the spatial and temporal intervals between any two visits of each user into appropriate weights and combines them in and adaptive manner.Then,it is proposed to integrate the spatialtemporal weight matrix into the multi-head self-attention network module,which enriches the personalized representation of user trajectories.3.A multi-relational heterogeneous and geography-enhanced graph neural network is proposed.In order to model the diverse interactions among users,POIs as well as geographical locations and explore geographical information features deeply,this model constructs a multivariate relationship heterogeneous graph of ”user-POI-region”.According to the different relations of nodes,different sampling sequences are obtained by defining sampling meta-paths.The embedding vectors of nodes under different relations are obtained by modeling with graph neural network.Then,a hierarchical self-attention network model of relationship-sequence is proposed to aggregate the node vectors under different relationships in an adaptive manner.Finally,the node vectors are added as auxiliary information into the trajectory of users.4.In view of the two proposed models mentioned above,this thesis makes comparison experiments,ablation experiments and sensitivity analysis experiments with the existing advanced models on widely used real-word datasets in related research.The experimental results show that the recommendation models proposed in this thesis have higher recommendation accuracy,which verifies the advancement and effectiveness of the two models.