Implementation Of Poi Recommendation Algorithm Based On Spatio-Temporal Information

POI recommendation is the core service of the location-based social networks,which aims to provide users with personalized and reliable visiting services.Based on users’ interaction records on social platforms,POI recommendation analyzes and mines uses’ potential preferences and pattern information in the data,and recommends locations that may be of interest to users.In recent years,the development of mobile Internet technology has driven users’ data growing exponentially and let data structures increasingly complex.This has led to an increasing demand for mining the valuable information in massive data,and thus,recommendation system,which learns user behavior patterns and provides differentiated data recommendation services,has become a hot research topic.This thesis is a research and implementation of POI recommendation algorithm based on spatio-temporal information.The existing research works lack the in-depth mining of combined information of time-space dimension and ignore the complex structural association existing between user records.In this thesis,three POI recommendation algorithms are proposed to address the personalization problem,the temporal pattern learning problem and the vector representation problem in recommendation research,respectively.The main work and innovative contributions of this thesis are summarized as follows.1.A graph convolutional recommendation algorithm is proposed for personalized POI recommendation based on temporal knowledge graph neighborhoods.This work introduces a temporal knowledge graph to the POI recommendation task and establishes a multi-entity heterogeneous temporal graph.The multi-level representation learning of the temporal information is able to improve the personalized recommendation performance effectively by considering the user’s dynamic,personalized,and geographic factor preferences in temporal and spatial graph spaces.The algorithm achieves a Top-1 accuracy of 25.7% in a publicly available dataset,which is higher than the accuracy of20.7% for other POI recommendation methods and improves by up to 24.1%.2.A multiple periodic geography convolution network is proposed for temporal pattern learning on POI recommendation.This work first analyzes and studies the periodic behavior of real user visiting data.Based on the phenomenological analysis,a neural network model is established on the framework of the temporal convolution network.The model innovatively combines binary and segment tree structures and temporal convolutional networks,and designs two new tree convolutional modules for multiple intervals and multiple spans.The algorithm achieves a Top-5 hit rate of 47%,which is higher than the 45% hit rate of other recommendation methods.The effectiveness of the algorithm in learning periodicity is also verified in real cases.3.A sequential POI recommendation algorithm with multi-scale vector mapping is proposed for the spatial vector representation problem.The work targets multi-fine grained spatial representation of spatial location relations,and can consider vector representation at three levels simultaneously: multi geo-scale level,local spatial and global spatial,which enables the algorithm to fully learn the directional relations and spatial layout during user movement,and improve the accuracy of user trajectory prediction.The prediction hit rate of algorithm Top-5 reaches 46.3%,which is higher than the hit rate of other methods with the same evaluation metrics of 44.1%.In summary,the POI recommendation algorithms proposed in this thesis investigate the in-depth combination of POI recommendation tasks with graph networks,convolutional networks,and vector representation methods in mining spatio-temporal information,and verify the effectiveness of the work on real datasets through extensive experiments,which provides an important reference value for POI recommendation research.