Research On The Next POI Recommendation Methods Based On Self-attention

With the rapid development of society and the popularization of positioning system technology and smart mobile devices,location-based social network(LBSN)services have gradually become an essential tool for people's daily social interaction and travel.In LBSN users can sign in to the points of interest and share their experiences about these visited points of interest on their mobile devices.Therefore,there are mass users' data in locationbased social networks.To obtain the next check-in points of interest that users will be interested in from these massive data and meet the personalized needs of users,the problem of points of interest recommendation has become an emerging research hotspot.In comparison with the traditional recommendation systems,the data on POI recommendations is sparse:(1)The check-in behavior of a single user is less,so it is more difficult to judge the user's next decision-making activity;(2)The user's check-in data samples are all positive samples,and the negative samples are lacking;(3)The interest preferences of the user will change over time;(4)The check-in points of interest of users have spatial aggregation within a certain period;(5)At the same time,we predict the task of the next POI based on the current user's location,which is continuous and influenced by a variety of contextual factors.In recent years,researchers have put forward some feasible solutions for the next POI recommendation,which aren't satisfactory for the next POI recommendation.In this work,we put forward a new solution for the next point of interest recommendation based on the existed research work.This thesis proposes a personalization framework to extract users' data efficiently for the above problems.The framework uses the self-attention mechanism to model the temporal features of the user's historical check-in sequence.To extract the user's dynamic preferences efficiently,we consider the temporal patterns of the user's historical trajectories in this thesis.This thesis integrates the user's Social relationships and spaces to solve the problem of the sparsity of user data.The work of this thesis is as follows:(1)The temporal characteristics of the user's historical check-in sequence are analyzed.We obtain the temporal relationship between the user's consecutive check-ins by fusing the relative time matrix of users' history track based on vanilla self-attention and complex and diverse preference features.Then we divide these preference features with time information into different time windows to obtain the user's short-term preferences.To reduce the search space of the model,we first perform data preprocessing on the candidate set,combining the geographic factors and popularity of POIs to filter out those POIs that are far from the user's current location and are not popular.Finally,we make corresponding recommendations for users according to the scores of the filtered candidate sets.(2)We fuse the user's spatiotemporal matrix to capture the temporal and spatial relationship of the user's historical trajectory based on the self-attention mechanism.We use collaborative filtering technology to model the preferences of users and their friends.We measure the user's and the friend's preferences through an attention mechanism to obtain the user's social influence.We utilize the MLP model to successfully fuse the information of users' social relations and spatio-temporal context,and propose a next POI recommendation model with users' social relation features and spatio-temporal context information.In this study,we demonstrate the effectiveness of our model through extensive experiments based on two large public datasets,namely Foursquare and Gowalla.