Research On Preference Processing Method Based On CUR Decomposition

Recommendation algorithm plays an increasingly important role in e-commerce platform.The quality of recommendation algorithm directly affects user experience,and then affects the income of e-commerce platform.Therefore,it is the focus of this paper to propose a good recommendation algorithm and optimize the proposed algo-rithm.Specifically,the recommendation algorithm needs to comprehensively analyze the historical user data and the real-time user data of e-commerce website,then analyze the user's preference characteristics,and finally recommend the interested products to users.Therefore,the analysis of user preferences is the key research content of the recommendation algorithm.Because the user's data is dynamic,it requires the rec-ommendation algorithm to be able to analyze the dynamic data.Because e-commerce platform stores massive user data,so the recommendation algorithm is facing big data problems,such as memory overflow,high computational complexity and so on.In view of the two problems facing the recommendation algorithm,We propose a two stage recommendation algorithm framework—CUR/C+RSVD.First,we use CUR to reduce data dimension,extract user's features preliminarily,and use RSVD to further extract user characteristics and recommend.It strives to improve the speed and accuracy of the extraction of preferences and the speed and accuracy of the recommendation.The main works of this paper are as follows:First of all,we use the CUR method to get a low-rank approximation of the original matrix A?user preferences for products?in order to extract the potential preferences of users and products.First,we use CUR to decompose the original matrix A into 3 low-dimensional matrices C?U and R.Thematrix contains the potential features of the item,and thematrix contains the potential characteristics of the user.Both thematrix and thematrix are made up of the real rows and columns in the original matrix,so the extracted user features and product features have good interpretability.Secondly,we use RSVD to prediction and recommendation.The main compu-tation cost in the original RSVD is to computeandrespectively.While replacingby CUR orin RSVD,we can reduce the computational cost from?8?²9))to??8²8?+?89?8?6)_?8??where?8 is the size of data subspace,8?and9?are the size of the input matrix,t is the number of iterations,6)_{?8?is the number of features}.Because CUR is explicitly expressed in terms of a small number of actual columns and actual rows of the original data matrix,the result of matrix decomposition has better interpretability.The advantage of what we devised CUR+RSVD and C+RSVD collaborative prediction approaches is that,they not only can deal with the large scale matrix rapidly,but also preserve the sparsity of the original matrix,more interestingly,CUR/C+RSVD has higher numerically stable and prediction accuracy.Finally,by analyzing the nature of the algorithm and analyzing the results of ex-periments,CUR can extract features of users and features of items at the same time.Compared with the traditional regularized singular value decomposition?RSVD?,sin-gular value decomposition?SVD?and other matrix decomposition methods,CUR matrix decomposition has the advantages of strong anti-interference ability,high com-pression rate,good interpretable ability,and high computational complexity and ac-curacy guaranteed by big data.CUR/C+RSVD can not only effectively deal with the problem of data sparseness,but also handle the low rank matrix scale.Compared with RSVD,it can achieve far better prediction accuracy and recommendation result?RMSE reduced about 20%,NDCG increased about 30%?,meanwhile save about 70%training time at the same dataset.