Research On Matrix Factorization Recommendation Algorithm Based On Improved SVD And Transfer Learning

With the rapid development of the Internet,the activities about people's production,life and entertainment are recorded more and more by data.In the field of popular electronic commerce,mining interests for personalized recommendation and analyzing online behavior of users(such as scoring,rating,browsing)are regarded as one of the most value directions.Nowadays,recommendation algorithms have merged many kinds of data,such as rating data,user item attribute information,social network information,tag reviews and even mobile location related data.Heterogeneous user data contained in a large number of user preference information,if we can fully tap the user's preferences,adopt appropriate model to filter noise,we will improve the accuracy of recommendation and user experience,which also produces enormous economic benefits,therefore recommendation has become a hotspot in academia and industry.Although much data provides a good data basis for recommendation algorithm,because of the large scale,multidimensional and even sparsity of data,the accuracy and real-time of algorithm still have much a lot of space to improve.This paper analyzes user's rating data,item attribute data and user's heterogeneous feedback data,finishes the work as follows:(1)When the matrix factorization algorithm solves the problem of data sparsity,it doesn't make full use of mutual information of preference information and user behavior score of the item,so we put forward the user UC-SVD algorithm,namely Singular Value Factorization Recommendation Algorithm Including User's Preference for Item attributes,considering the item attribute and user rating behavior,constructs a user preference matrix for item attribute,representing a user preference for certain types of item,and the item attribute and user preferences on the item properties are added into the matrix factorization model to make up the problem of rating data to some extent.(2)Aiming at the accuracy of recommendation algorithm,we put forward the HFBT algorithm combining explicit rating data and implicit feedback data and take two steps to process explicit rating data and user implicit feedback data.Firstly,by drawing lessons from SVD++ algorithm we introduce the user preference factor and non-user preference factor and two factors are integrated to matrix factorization model;the second way is that we learn from the transfer learning to regard the matrix factorization including the implicit feedback as transferring auxiliary resource field,the item feature factor is migrated to aiming rating field and the influence of user feature factor is kept,which can achieve better recommendation accuracy.(3)Aiming at the problem of more parameters and low accuracy about the HFBT algorithm,we parallelize the HFBT algorithm proposed in this paper on the Spark,effectively improve the efficiency and scalability of the algorithm.In the real Spark cluster environment,the proposed algorithm HFBT is compared with other recommended algorithms,and the accuracy of the recommendation is verified under three datasets.The results show that the recommendation algorithm proposed in this paper not only improves the accuracy of recommendation,but also has good scalability in large-scale datasets.