Research And Implementation Of POI Recommendation Algorithm Based On Deep Learning

With the continuous advancement of science and technology,location-based social networks have also been greatly developed.POInt-of-Interest(POI)recommendation has received a lot of attention as an important task for location-based social networks.The purpose of a POI recommendation is to recommend POIs of possible interest to a given user based on the user's historical check-in history.In recent years,deep learning has made major breakthroughs in various fields of artificial intelligence,leading to a boom in the field of artificial intelligence reform and bringing new opportunities to the research of recommendation system.Deep learning has also recently been widely used in POI recommendation systems,which will dramatically change the architecture of traditional POI recommendation,opening up new opportunities to further improve the user experience.In POI recommendation,data sparsity has always been a vital problem perplexing POI recommendation.Due to the sparse check-in history of users,it is difficult for POI recommendation algorithm to directly understand users' preferences through the check-in history of users.In the existing studies,the problem of data sparse is generally alleviated by adding various ancillary information,such as the check-in time of users',the geographical location of POI and the social relationship of users.However,they do not take full advantage of negative user feedback and category information of the POI.In addition,now the transportation is gradually convenient,people may often leave their usual familiar areas for a variety of reasons to visit relatively unfamiliar areas.In unfamiliar areas,the user has fewer activity records,which makes the data sparsity problem more serious.The common POI recommendation algorithm does not take this scenario into account.However,with the progress of society and the convenience of transportation,such demand becomes more and more common,which puts forward new requirements for the recommendation algorithm of POI to provide high-quality POI recommendation when users go to unfamiliar areas.In view of the above two POI recommendations,this paper studies from the following two aspects:(1)We propose a recommendation model based on Bayesian Personalized Ranking and graph neural network for POI recommendation.Specifically,we first use Bayesian Personalized Ranking to generate a representation of user preferences and POIs.POI category information and geographical location information are integrated in the negative sample sampling.Then,graph neural network and long and short memory network(LSTM)were used to obtain sequence features based on geographic location and POI category,respectively.Finally,the features extracted from the two networks are fused together to obtain the final probability of candidate POI.And verified it on Foursquare,a dataset of real cities.Experimental results show that the proposed algorithm can effectively alleviate data sparsity and cold start problems,and get better recommendation performance.(2)Aiming at this new scenario of POI recommendation in previously unvisited locations,this paper proposes a remote POI recommendation model based on popularity characteristics analysis and social network.In order to alleviate the data sparsity and cold start problems faced by this problem,we integrated two kinds of auxiliary information into the model,namely social information and popularity information.When integrating social information,we take into account that users' trust in their friends is influenced by the degree to which their interests are similar.When incorporating the popularity information for POI,we took into account that the impact of popularity on user check-in varies over time and distance from the POI.The results were verified on Yelp,a real urban dataset.The experimental results show that the algorithm proposed in this paper can effectively alleviate the data sparsity problem in the recommendation in previously unvisited locations,so that the recommendation results can achieve a high precision and recall rate.