The Research Of POI Recommendation Algorithm Based On Deep Learning And Its Improvement

With the booming of the Location-Based Social Network(LBSN),the Point-of-interest(POI)recommendation service plays an indispensable role in our daily life.So it becomes an important part in the research of the recommendation service.In the research of POI recommendation service,it faces two problems.One is the data sparse of the rating of POIs which is rated by user.The other is the representation of the feature vectors of the users and POIs.To alleviate the problem of data sparse,the tradictional recommendation algorithm based collaborative filtering(CF)works but without good performance.With the booming of deep learning neural network(DNN)technology,more and more researchers apply it on recommendation service these years for it can increase or decrease the dimensions of a data's represntaion.And the technology of users and POIs' representation gets a new chance with the widely use of probability matrix factorization(PMF)algorithm and knowledge graph.To overcome the problem of data sparse and get better users and POIs' feature vectors representation,this paper worked as follow:(1)To review the-state-of-art recommendation algorithm,evaluation method and frequently-used datasets.Then proposed the framework of POI recommendation based on LBSN.These provide the basic tools for the research work of POI recommendation.(2)To archieve better users and POIs' representation of feature vectors through the PMF algorithm fused various information.Then proposed a recommendation model named TGSSMF with CF recommendation algorithm.Compared with a POI recommendation model based on DNN in which is used to learn the iteraction of a user and a POI,the experiment shows that the model based on DNN performs better than the recommendation algorithm based on matrix factorization with single information.But it loss to the TGSS-MF model which fused with various information,and construct by the PMF algorithm and CF algorithm.(3)To proposed a POI recommendation model named AKG-DNNRec which construct by the knowledge graph and DNN.This model alleviate the information lost during the walk on the knowledge graph through the DNN.So this model provides a better users and POIs representation.By the analysis of the experiment,it proves that the model gets better performance in recommendation compared with the recommendation algorithm based on knowledge graph embedding and DNN respectively.