Application And Research Of Recommendation System Based On SVD

Recommendation system is a product of the rapid development of the Internet,and it has been become more and more important in our daily lives,work and study.Nowadays,the recommendation system has been developed rapidly in the fields of e-commerce,film,social,music,and so on,and the application and research of the recommendation system had always the research focus at home and abroad.Recommendation algorithm and big data which algorithms relied on are the hardcore of a recommendation system.SVD technology can be used to study for the binary relation rating data of user-item and the ternary relation weights data of user-item-tag in recommendation system,and it is currently the key and effective algorithm which can simultaneously deal with the two kinds of data.However,in pace with the increasing size of data should be processed,the low computational efficiency and recommend accuracy become the key problem of the system.This paper aimed at the problem of the low computational efficiency and the not ideal recommend accuracy of the SVD technology,and made a deep research on the performance of the low and high order singular value decomposition algorithm,the main works include the following three parts:1.Firstly,the paper researched on the performance of the LFM,Bias SVD,and SVD++ recommendation algorithms which were all improved base on the SVD basic algorithm.Among the three,LFM decomposed the high dimensional matrix into two low dimensional characteristic matrices of users and items.Bias SVD was added the bias information of user and item into the model based on the LFM.SVD++ was added the implicit information into the Bias SVD model.After using the theory and experiment to analyze and compare the performances of the three models,it can be found that SVD++ was the best and LFM is the worst in terms of accuracy,while LFM was the highest and SVD++ is the lowest in terms of computational efficiency.2.Secondly,the paper studied deeply in the low efficiency of SVD++,which caused by the high computational complexity.When dealt with the binary relation rating data of user-item using SVD++ algorithm,although it can get higher accuracy,but it considered many factors and the computational complexity was large,while the computational efficiency was very low.Through analyzing the theoretical model found that when using the gradient descent method to train the objective function of prediction model,the learning rate function directly affected the iteration number and the convergence speed after training.The traditional learning rate function with exponential function had many defects,such as decreased slowly,too many iterations and so on.So,the paper proposed a new learning rate function to train the SVD + + model,the learning rate function had the features of low initial value,falling rapidly in themedium-term and small value with slow changes towards the end of the computation.The experiment result revealed that the method can not only improve the computation efficiency of SVD++ model but also guarantee the accuracy prediction under the premise of using gradient descent method to train the algorithm model.3.At last,this paper studied the recommendation based on the ternary relational data of user-item-tag.In recommendation system,use-item-tag data often suffers from tag redundancy.Therefore,if make full use of the characteristics to find the relevance between tags,it will be useful to improve the efficiency of prediction.So,this paper proposed a HOSVD recommendation algorithm based on Apriori algorithm to recombine the tags.The algorithm firstly adopted Apriori algorithm to pre-treat the original tag data,firstly search the frequent tag item set,then set the item set as a new tag,and numbered the new tags,so the new user-item-tag data generated,and then the HOSVD algorithm was adopted to process the new data.Through the experiment,it showed that the performance of the proposed method is obviously improved.