Research On Successive POI Recommendation Method Based On User Inherent And Contextual Preference Awareness Fusion

With the development of mobile Internet and smart devices,Location-based Social Network(LBSN)has been rapidly developed,such as Foursquare,Yelp,Places and other online social platforms,which attract more and more users to record and share their personal lives using mobile devices.The availability of large amount of user interaction data makes it possible to provide more personalized and accurate recommendation services,and the problem of successive Point-of-Interest(POI)recommendation has received extensive attention from academia and industry.Successive POI recommendation refers to the recommendation of POI that may be of interest to users in the future period based on their historical check-in records.However,in mobile scenarios,multiple influencing factors such as the diversity of user preferences,the variability of user behaviors,the dynamism of spatio-temporal contexts and the high sparsity of check-in data bring great challenges to the user interest capture capability and response real-time requirements of POI recommendation systems.Based on the above challenges,this thesis focuses on the successive POI recommendation system with dynamic user preference perception,and conducts research on solving the complex interest perception fusion,data sparsity problem handling,and online recommendation performance improvement of successive POI recommendation,etc.The main contents include:(1)To address the difficulty of preference capture caused by context-dependent dynamic preference,user interest diversity and drift in mobile environment,this thesis proposes the inherent preference module is trained based on users’ global historical behavior data to achieve inherent interest representation,while the contextual preference module is trained based on context-sensitive data selected on-the-fly to achieve dynamic interest perception,and by jointly training users’ inherent and contextual preferences,the fusion generates users’ preference representation in mobile scenarios and carries out successive POI recommendation.(2)For the online response requirement of successive POI recommendation in mobile environment,based on the fast growing massive POI training and online data,this thesis proposes to introduce the R-tree based POI index structure and the improved IR2-tree based trajectory index structure for POI candidate set retrieval and context-sensitive training data retrieval stages,respectively,to improve the real-time response capability of the recommendation system.(3)To address the problem of highly sparse check-in data,this thesis proposes a meta-optimization model based on few-shot recommendations,which integrates the tasks of user inherent preferences and contextual preferences in meta-learning,and optimizes the recommendation model based on preference-aware fusion through the meta-optimization framework,which can quickly adapt to users with fewer check-in records and improve the recommendation effect for such users to a certain extent.In order to verify the effectiveness of the attention network based on user preference-aware fusion,this thesis conducts a large number of experiments on three real datasets,NYC,TKY and Weeplaces,and the experimental results show that the proposed method has significantly improved the recommendation effect compared with other methods,which verifies the effectiveness of the attention network method.In order to verify the effectiveness of the metaoptimization model based on few-shot recommendations,experiments are conducted on two datasets,NYC and TKY,and the experimental results show that the meta-optimization model proposed in this thesis has better recommendation performance compared with other methods.