The Design And Implementation Of Recommender System Based On Learning To Rank

In recent years,with the rapid growth of location-based social network,Yelp,Foursquare and Meituan services emerge as the times require.Point-of-Interest(POI)recommendation has become one of the most cutting-edge fields.In location-based social network,user can check-in and find more attractive POIs based on the recommendation.However,the extremely sparse user check-in matrix poses some challenges for POI recommendation.In addition,different types of contextual information are useful complements to the user preferences and can be fully applied to the POI recommendation.In order to solve these challenges,this paper studies and designs an POI recommendation system based on the learning to rank.The main work is as follows:(1)A POI recommendation algorithm based on sentiment analysis is proposed,which is also a point-wise ranking algorithm.The algorithm takes the contextual information of POIs,neighborhood,classified friends and popularity into account.Through data analysis and observation,we study the relationship between neighborhood and classified friends,and find that their neighbors tend to have weak correlation to the POIs,and a specific category friend is more similar than a generalized friend.In addition,according to the emotional preference analysis of the users'comments,the user's real preference is deduced,thus the implicit feedback is converted to 1-5 score.(2)We propose a POI recommendation algorithm based on ranking weighted matrix factorization(RankWFM)for behaviour data which missing comments.Sepcifaically,we aim to maximize the probability of the location in the recommendation list,and redesign the objective function.The unchecked places also help the algorithm learning,which will help to alleviate the sparsity problem.In addition,the geographical location and business classification can be prori weighted knowledge,which is integrated into the weight matrix decomposition.Experimental results show that,compared with no ranking optimization algorithms such as Geo-MF and PMF,Rank WFM algorithm has greatly improved the recall and precission measurement.(3)We propose a ranking POI recommendation based on deep forest.The algorithm takes NDCG as ranking measurement,and using deep forest to form deep network.Each layer of the newwork aims at fitting Lambda residuals,and then learns the implicit preferences of users and interest points.In addition,the algorithm can automatically construct more grained features and learning layer,which obtain a stronger implicit learning ability.We experiment on the foursquare dataset,and evaluate the performance of NDCG,which is 8%heigher than the LambdaMart algorithm on the NDCG measurement.(4)An POI recommendation system based on learning to rank is designed and implemented.The system includes a feature selection layer,the candidate set trigger layer,fusion filtering layer and reorder layers.This system not only helps local residents or tourists to explore an POI,but also to help businesses find and attract potential tourists.