Research On Recommendation Model Based On Ensemble Learning

With the development of Internet technology in recent years,users face thousands of products every day.However,it is a difficult problem to choose the useful from products.As an independent research direction,recommendation system is favored by researchers and industry,it provides users with personalized services to solve the problem of information overload.Although recommendation system has developed rapidly in the past decades,there are still many performance problems to be solved urgently.Matrix decomposition based recommendation algorithm has been a hotspot in research because of its simple idea and high scalability.Futhermore,text information,social information and graph information can be combined to improve performance.So the basic matrix decomposition algorithm is very important.Existing probability-based matrix decomposition representates the original score matrix with a low-dimensional latent factor matrix.Although the idea is simple,many information hidden in the scoring is ignored.Therefore,the performance is poor.However,joint-clustering based methods have provided new ways of thinking.In this thesis,the matrix approximation algorithm based on joint clustering is studied and discussed in depth.The research work of this paper is summarized as follows:Firstly,based on the idea of decomposing the original score matrix into two low-rank users and representing the potential factor matrix of items,we use maximum likelihood function to update the latent factor matrix of user items iteratively.In this paper,an Adaptive Weight Matrix Factorization model is proposed based on the traditional probabilistic matrix factorization model.The model AWMF proposed in this paper learns to predict the correct weights by introducing Laplacian distribution.The predictive accuracy is high while the corresponding weight is large for users with accurate predictionsSecondly,the boosting theory is deeply studied,a matrix approximate recommendation algorithm based on boosting framework is proposed by the idea of adding model and forward stage wise algorithm.We use boosting framework to integrate the score prediction matrix with multi perspective.Based on boosting theory,the proposed AWMF model is a regression problem with square error loss.We use the gradient boosting algorithm,in other words,the approximation method of steepest descent method,to fit the regression problem with the negative gradient of loss function.Thirdly,a matrix approximate recommendation model combined with global and local information is proposed.Since the traditional matrix approximation method mainly considers the overall structural information of the scoring matrix,it does not consider the local correlation of the user.This paper proposes a matrix approximation algorithm based on joint clustering.Inspired by clustering the users of the original scoring matrix and item clustering,we iteratively generated multiple templates to predict the final score.Based on user's scoring information,scoring matrix is divided into local correlation scoring blocks by clustering.Based on the boosting framework proposed above,a matrix approximate recommendation model of joint clustering is constructed to predict the vacancy from the original scoring matrix.This method improves the performance of the proposed algorithm by effective fusion clustering,and has better robustness and stability.Finally,in order to verify the recommended performance of the two matrix-based matrix approximation recommendation models constructed in this paper,we compare the test results on the recommended standard dataset by cross validation.The experimental results show that the proposed matrix approximation algorithm based on boosting framework has better recommendation performance.The two matrix approximation algorithms obtained according to different probability maps relatively improve the recommendation performance.Simultaneously,the matrix approximate recommendation model proposed in this paper,which integrated global and local information,has good experimental performance on sparse and dense datasets.This indicates the joint clustering model based on matrix approximation is robustness and stability.