Research On Recommendation Algorithm Based On Deep Learning

With the rapid development of computer networks and mobile Internet,e-commerce and location-based social networks(LBSN)have also developed rapidly,such as Amazon,Alibaba,JD,Facebook,Foursquare,Gowalla,etc.At the same time,the problem of information overload also Increasingly serious,how to mine valuable information for users from massive data and reduce information redundancy has become a hot issue in industry and research.This is also a problem to be solved by the recommendation system.The recommendation system is mainly faced with the following problems: data sparseness.Due to the small amount of interaction data between users and projects,the user-item interaction matrix is too sparse,making the recommendation system unable to make accurate recommendations by using the user's sparse history.Then the second problem is cold start.For new users,since there is no historical record,the recommendation system will naturally not be able to use its historical record to make recommendations.The cold start problem can be seen as the extreme of data sparsity.In the real world,users and items have rich contextual information,such as the user's social relationship,work,age,geographic location of the item,tags,attributes,prices,etc.,and more and more people begin to use auxiliary information to alleviate These two problems.At the same time,since context information belongs to heterogeneous data,how to integrate these heterogeneous context information has also become a major challenge.On the other hand,different users have different check-in distributions,and the existing location-based POI recommendation algorithm fails to make full use of the user's checkin distribution characteristics and regional characteristics.Traditional recommendation algorithms mainly use historical interaction data of users for modeling,such as matrix factorization algorithm,one of the most classic recommendation algorithms.The dot product of user vector and item vector is used to predict the probability.However,the use of simple linear relationships for modeling cannot reflect the complex nonlinear relationships in the real world.With the great success of deep learning in natural language processing,image processing,speech recognition and other fields,more and more people are applying deep learning to recommendation systems.In this article,we will also use deep learning techniques to solve.In view of the problems in the above recommendation algorithm and in order to recommend more valuable products to users,this article mainly studies from the following three aspects:(1)This article propose a new model(SATCo NN)based on self-attention mechanism to fuse user similarity and item similarity.SATCo NN is mainly based on the recurrent neural network RNN module,and uses the self-attention mechanism to obtain the weight of each item in the user's purchase history from different semantics.Drawing on the idea of style transfer,we use the Gram matrix to model the user's shopping style,and we use the Maxpooling technology to extract the user's shopping style.(2)This article can fuse a variety of auxiliary information.In this article,we propose an unsupervised user check-in distribution feature extractor,CD-Ex,which can unsupervised learn user check-in distribution features.We also propose user-based Collaborative filter trees and item-based collaborative filter trees,we borrow the idea of message passing to learn the deep representation of users and items.In these two modules,we use the multi-head attention and original attention mechanism to learn the representation of users and items.Regarding the user's social relationship,we use the user's friends to assign weights and aggregate the user's friends in the user-based collaborative filtering tree.(1)In this article,we can combine regions and POIs to make recommendations.We meshed the user check-in areas and through the statistics of the check-in probability in each area obtaining three different granularity check-in distribution matrices,namely user check-in distribution matrix,friend circle check-in distribution matrix,and regional heat matrix.The three matrices are weighted and summed to obtain the user's check-in distribution vector through the convolutional neural network.We also propose four modules,the user preference module,the check-in distribution module,the item heat module,the region indirect module.The item heat module can alleviate the cold start problem.