Research On POI And Route Recommendation Algorithm Based On Deep Learning

With the progress of society,the mobile Internet is also showing a trend of vigorous development.Online e-commerce platforms and location-based social networks(LBSN)have also developed rapidly,such as Taobao,JD,Facebook,Twitter,Foursquare,Gowalla,Yelp,etc.At the same time,issues such as information explosion have also received increasing attention.How to extract information that is meaningful to users from the explosive growth of Internet massive data,while filtering irrelevant information as much as possible,has become an important research problem in the industry and research circles.This is also the problem to be solved by the recommendation system.At this stage,the recommendation system mainly faces the following problems.First,the issue of data sparseness,because the interaction data between users and items that we can collect is tiny,resulting in the built-up user-item interaction matrix is too sparse,in this case,under the circumstances,it is extremely difficult to use the sparse user history records to make more accurate recommendations through the recommendation system.The second cold start problem is that for new users who have just registered,because they have no historical interaction records.It is not feasible for the recommendation system to use their historical interaction records to make recommendations.Nowadays,people are beginning to use the ancillary information of users and projects(such as the user's social relationships,changes in interests,career directions,project location,classification,rating,price,etc.)to alleviate data sparseness and cold start problems.Traditional recommendation algorithms,such as matrix factorization algorithms,often just use the dot product of the user vector and the item vector to predict the probability.But this kind of modeling only through linear relationship does not reflect the complex nonlinear relationship in the real world.As deep learning technology has achieved amazing results in hot areas such as speech recognition and natural language processing,this article will also use deep learning techniques to solve our problems.In response to the problems mentioned above and to provide users with more accurate recommendations,this article mainly researches the following three aspects:(1)Research the point-of-interest recommendation algorithm based on self-attention mechanism.This paper proposes a self-attention mechanism-based model(SSANet)to integrate user similarity and item similarity.SSANet uses multi-hot to model the sequence data in the user history to obtain the user embedding representation;then model the POI ancillary information to obtain the ancillary information embedding representation;then the fusion output of the user embedding and the ancillary information embedding uses self-attention mechanism to learn the representation of user history from different aspects.Experiments show that our proposed model has achieved a significant performance improvement compared to other models.(2)Research the point of interest recommendation algorithm based on graph attention mechanism.This paper proposes a model based on the graph attention mechanism(SGANet),which can learn user preferences unsupervised,and at the same time,based on the session window to further feature extraction of the user's regional history check-in data,and introduces graph attention mechanism captures the user's preference from POI pairs and regional POIs within the model and further learns the longterm and short-term preferences of users through the integration of POI auxiliary information and the use of recurrent neural network GRU.A series of experiments show that our proposed model has achieved significant performance improvement over other models.(3)Research route recommendation algorithm based on generative confrontation network.This paper proposes a model(RGANet)based on the generative confrontation network mechanism.Since the user's route feature is sequence data,to extract the user's route feature,this article uses GRU to fetch.This article proposes three modules,the data integration module,the route generation module,and the route recommendation module.In the generator part,a route is generated according to limited conditions,and GRU is used to extract route features.In the discriminator part,the structured POI sequence of the user also uses GRU to extract route features,and then calculates the similarity of the route features,so that the generator part generates routes with higher user satisfaction based on the feedback from the discriminator part.Experiments show that the performance of our proposed model is better than other models.