Research On Recommendation Algorithm Of Matrix Factorization Based On Mining User And Item Information

In recent years,recommender systems have developed rapidly,as one of the commonly used collaborative filtering recommendation algorithm,matrix factorization model has been used widely,this thesis mainly focuses on how to effectively use matrix factorization as the basic model to predict ratings.We combine ensemble learning with matrix factorization to extremely solve the different effect of different type of data in predicting ratings.Recently,the personalized recommendation in Internet has been more and more important,we use the SVD++ model which combines the users and items themselves preference characteristics to get personalized recommendation,in order to better implement personalized recommendation rapidly and efficiently,we combine SVD++ with Pearson coefficient and association rules to gain more deeper personalized recommendation to users,which could improve the accuracy and efficiency of recommendation.In the second chapter,we conduct a thorough research to the existing model of matrix factorization,and combine ensemble learning with matrix factorization to further mining the inner imbalanced character of data.By ensembling matrix factorization model,the precision of recommendation algorithm has been further improved,ensemble learning can avoid imbalance and redundancy of data to a certain extent.From the results,the ensemble matrix factorization model improved the prediction accuracy than traditional matrix factorization model to some extent.In the third chapter,after summarizing the algorithm in the second chapter,we consider that,although ensemble matrix factorization model improved the prediction accuracy than traditional matrix factorization model to some extent in the second chapter,the matrix factorization model couldn't take inner association information of users and items themselves into account,and after applied these implied inner information in collaborative filtering recommendation algorithm can largely improve the recommendation accuracy.SVD++ is also a kind of matrix factorization model,it introduces the implicit feedback,using the history of users' rating datum.But,SVD++ algorithm takes all of the history rating data of users into consideration,which was high complexity,time consuming and impractical in actual application.So,we mining the potential information between items,association rules is an effective method to mine inner relation between datum.In this chapter,we take each user's labeled items as an entry,since these items were labeled together in this entry,so there exists a certain correlation between them to a great extent.Therefore,we use the association rules to deal with these entries and mining the associated items in entries,during training the model,we use these associated items to update the characteristic matrices of users and items to realize effective acceleration.In the fourth chapter,we mine the similar items set of item from another way.We use the Pearson coefficient to mine the similar items set of every item in this chapter,in the third chapter,association rules take each user's labeled items as an entry from the points of users.Here,from the points of items,we take all ratings of every item given by users as a vector to calculate the similarity between items,and find out the similar items set of every item.In the process of training model,we use the similar items set of item to replace the items set which marked by user to realize effective acceleration.