Research On Next POI Recommendation Method Based On Multi-feature And Graph Representation Learning

Location-based Social Networks(LBSNs)provide a social platform in which users can check in to places,and share ratings,photos,and comments.With the proliferation of smartphones and the growing user demand,massive check-in data has given rise to a new research direction,which is the recommendation of the next Point-of-Interest(POI).Next POI recommendation can suggest the next visiting location that users might be interested in based on their historical check-in records and contextual information,thereby greatly improving the user experience.Recently,deep learning has provided an effective technical means for solving the next POI recommendation problem,which has achieved some research results in POI recommendation.However,due to the complexity of the problem scenario and the variability of user needs,current methods still face many challenges: on the one hand,current methods can usually only model a single behavior pattern and cannot simultaneously learn multiple behavior patterns of users;on the other hand,existing context-aware models mostly only model relationships between single pair of context instances and lack the ability to learn relationship among multiple context instances.To address these limitations,this thesis first proposes a self-supervised learning model based on graph,which can obtain auxiliary supervision signals from different behavioral patterns to achieve multi-behavior pattern learning and improve recommendation performance.Furthermore,a multi-relationship heterogeneous graph convolution method is proposed,which can stably fuse feature representations of different data relationships in the graph convolution process,thereby improving the robustness and recommendation accuracy of the model.The main contents of this thesis are as follows:(1)To address the problem of modeling personalized behavior patterns for POI recommendation,a graph self-supervised learning model is proposed.This model introduces the concept of implicit behavior patterns based on real-world LBSN check-in data,and utilizes two graph data augmentation operations to generate enhanced trajectory subgraphs for modeling implicit behavior patterns.Furthermore,a graph preference representation encoder is used to learn the high-level representation of the enhanced trajectory graph,and an asymmetric graph contrastive learning architecture is employed to maximize the consistency of the embedded representation of the enhanced trajectory subgraph based on positive samples,thereby learning the personalized behavior patterns of users.Finally,a multi-feature self-attention mechanism is used to learn users’ shortterm dynamic preferences and make recommendations based on the high-level representation of users’ trajectory graphs.Experimental results demonstrate the effectiveness of the proposed graph self-supervised model on three real-world datasets.(2)To address the problem of learning the relationships between multiple contextual feature information,a next POI recommendation model based on multi-relation heterogeneous graph convolutional network is proposed.Firstly,a multi-relation heterogeneous graph is constructed using existing contextual information and check-in behavior,including explicit relations between users-users,users-POIs,and POIs-POIs,as well as implicit relations guided by POI categories.Secondly,a heterogeneous graph convolutional network based on attention mechanism is designed to fuse the embedding features of multiple relations through attention mechanism,to ensure the stability of the graph convolution on the heterogeneous graph.Finally,an attention mechanism based on location offset embedding is proposed to obtain the feature representation of users’ dynamic decision-making information for predicting their next behavior.Experimental results show that this model outperforms baseline models on two datasets with different sparsity,and its robustness is verified on sparse datasets.