Multi-angle POI Sequence Recommendation Based On Self-attention Mechanism

With the popularity of social networks in recent years,the rapid development of locationbased social networks(LBSN)applications has been promoted.Coupled with the maturity of geolocation technology,the personalized recommendation of Point-of-Interest(POI)has become a popular branch in recommendation system direction.POI recommendation can not only discover users’ personalized preferences in the massive historical check-in data,but also help merchants provide better services to users according to users’ behavior habits.The POI recommendation system mainly recommends points of interest at the next moment according to the user’s personalized preferences.However,in the current research on POI recommendation,it is difficult to make full use of contextual information due to the sparseness of user check-in datasets.In addition,past research often pays the same attention to the amount of past information,and it is impossible to determine which part of the past checkin sequence is more significant for POI recommendation,so the efficiency and performance of conventional POI recommendation models are poor.Therefore,how to extract the latent features of users from the massive check-in data,fully excavate the context information,for instance,time,space,POI category and comment semantics,then integrate them into the POI model has become a difficult problem to improve the accuracy of POI recommendation.This paper proposes a multi-angle POI recommendation model(MA-PRM)that integrates spatiotemporal and emotional information and a multi-angle POI recommendation model(SAMA-PRM)based on self-attention mechanism,respectively.The main work contributions of my thesis are as follows:(1)My thesis proposes a POI sequence recommendation model MA-PRM that integrates spatiotemporal and emotional information.First,the time interval and geographic location interval of POI check-in are considered,and spatiotemporal features are extracted based on a symmetric matrix.Then,the POI category features are extracted based on inverse geocoding.Finally,based on the introduction of the public sentiment dictionary,the sentiment features in the user comment information are extracted by using the CBOW model.The embedding vector representation of the spatiotemporal,POI and emotional features of the user accessing a specific POI at a specific time is obtained,and the spliced POI embedding vector is fully mined by using the gated unit structure.The MA-PRM model can extract POI features from multiple perspectives,it effectively relieves the POI data sparsity and cold start problems,and bestially promote the performance of the POI recommendation model.(2)This thesis introduces a self-attention mechanism based on the MA-PRM,and proposes a SAMA-PRM model.The SAMA-PRM uses this mechanism to assign different weights to the historical POI check-in sequence,thereby judging which part of the historical check-in sequence has a greater impact on the model predicting the next check-in point.It can autonomously learn users’ long-term and short-term POI preferences and related relationships,and achieve better performance than the baseline model.