Next POI Recommendation Algorithm Based On Graph Representation Learning

Under the boom of Location-based Social Networks(LBSN),a large amount of check-in data has been generated.To better utilize this data and provide better services to LBSN users,the academic community has turned its attention to the task of recommending the next point of interest and has used graph representation learning to enhance the effectiveness of recommendations.However,modeling user spatial behavior and social relationships in LBSN graph representation learning is still a challenge due to the diverse types of contextual factors such as geographical influence,user social relationships,and time information.Likewise,the task of recommending the next point of interest also faces the problem of incomplete modeling of user long-and short-term preferences.To address these issues,this thesis focuses on LBSN graph representation learning and next point of interest recommendation,using user check-in data in LBSN as the research object.The study follows two main research directions:1.Graph representation learning on location-based social networks.To address the problem of insufficient modeling of user spatial behavior and social relationships,this thesis investigates the MCGL model,a graph representation learning model that combines multiple contextual information in LBSN.First,this model designs a weight strategy for the LBSN graph based on the analysis of LBSN data and constructs a fused LBSN graph that integrates contextual information.Then,this model utilizes graph embedding technology to learn the representation of the LBSN graph.Experimental results on two public datasets demonstrate that MCGL outperforms existing models in downstream tasks of LBSN graph representation learning regarding recommendation accuracy.2.Next point of interest recommendation based on MCGL.To address the incomplete modeling of user long and short-term preferences,this thesis proposes a long and short-term preference modeling model based on MCGL.First,this model obtains the embedding vector representations of users,points of interest,and time periods in the LBSN dataset using the MCGL model.Then,this model leverages Transformer encoder and non-local operations to model the long-term preference of users that fuses spatiotemporal information and designs a Transformer encoder based on geographic expansion to fully model the short-term preference of users.Finally,experiments on two public datasets demonstrate that the proposed model achieves better performance in Recall and NDCG metrics compared to existing models.