Research On Point-of-Interest Recommendation Model Based On Spatio-Temporal Context Awareness

With the rapid development of wearable devices,location-based social networks(LBSN)have received increasing attention.POI is recommended as a research hotspot of LBSN.It provides users with personalized services through the analysis and mining of massive check-in data.This not only allows the platform to better understand the target audience,but also brings great convenience to users,and is therefore widely used in various fields.Compared with the traditional recommendation system,POI recommendation has more technical challenges.On the one hand,the frequency of user check-in is too low,resulting in data that is too sparse,and it cannot reflect the user's dynamic movement behavior,so most association analysis algorithms are not effective.On the other hand,the user's preference is affected by different factors,and it has the characteristics of complexity and difference.Whether it can comprehensively consider the diversity of user preferences has become the key to achieve POI recommendation.In order to better realize the personalized recommendation of POI,this paper adopts the method of deep neural network to regard POI recommendation as the recommendation problem of sequence association to better capture the timing relationship of different user behaviors.First,by analyzing the impact of time,space and context information on user behavior,the heterogeneous data signed in by the user history is represented by the low-dimensional representation of its potential vector,and the sequence is fused and embedded into the encoded GRU network.Then,under the sequence-to-sequence framework,GRU is applied to the POI personalized recommendation model.In this framework,because the variability of the input sequence length is easy to cause information loss,in order to solve the situation that the sequence is too long and information loss occurs,a POI recommendation model based on spatiotemporal attention,STAPR),so that GRU can be better applied to POI recommended scenarios.In the STAPR model,through the global and local multi-layer attention,the encoder part based on spatial influence is constructed from different levels.Similarly,during the decoding and reconstruction process,the user's check-in information is modeled by introducing a time attention mechanism,which not only solves the problem of performance degradation over time,but also recommends users more effectively.Finally,in order to verify the effectiveness of the model,considering the possible limitations of the POI model by different densities,this paper verifies the data in four real-in-signature datasets with different densities.The experimental results show that the STAPR model has better effects on the data sets with different densities compared with the existing methods,and the higher the density of the data set,the better the performance,which reflects the generalization performance of the STAPR model.