Research On Recommendation Algorithm Based On Matrix Factorization

In recent years,with the rapid development of Internet technology and the widespread popularity of mobile devices,the amount of information in the Internet has exploded,causing the problem of "information overl oad".In order to alleviate this problem,the personalized recommendation system leverages user information,item information and other related content to predict the utility or relevance of a specific item,and then provides personalized recommendations to the user.This method actively explores users'preferences,helps users discover interesting content,enhances user experience,and alleviates the problem of "information over-load".It has been successfully applied to e-commerce,music,video and other fields.At present,in order to further improve the performance of the recommendation system,the research on the personalized recommendation algorithm is deepening,and recommendation algorithms in different scenarios have been successively proposed.Among all recommendation algorithms,matrix factorization has become a research hotspot in the field of recommendation because of its high r-ecommendation accuracy,strong scalability and fast running speed.This thesis analyses some problems existing in traditional matrix factorization recommendation algorithms,and proposes two improved matrix factorization based recommendation algorithms to enhance the recommendation performance.The main work and contributions of this thesis are as follows:(1)This thesis proposes a matrix factorization recommendation algorithm based on rating reliability.In the traditional matrix factorization algorithm,it is generally believed that the user's ratings are authentic and reliable,and each user will make an objective evaluation according to his own situation.However,in the actual situation,the final evaluation made by the user is affected by many factors and cannot fully reflect the user's real preference.In view of above problems,this thesis first proposes the definition of the reliability of rating data,measures the reliability of each user's rating data,and then proposes a matrix factorization recommendation algorithm based on rating reliability,which makes the high-reliability ratings get more attention in model training and makes the matrix factorization algorithm grasp the whole of users and items more accurately.The experimental results show that the matrix factorization recommendation algorithm based on rating reliability can further mine the feature information of users and items from the massive interactive data,and improve the overall performance of the recommendation system.(2)This thesis proposes a matrix factorization recommendation algorithm based on clustering relationship.In recommendation system,the number of users and items is very large,so global matrix factorization easily ignores the association between strong similar users or strong similar items.At present,clustering-based matrix factorization recommendation algorithms can mine the association between strong similar users or items,but they don't effectively deal with the local and global associations between users and items,which aggravates the sparsity of data.In view of the above problem,this thesis first proposes an improved clustering algorithm to find similar users and similar items more accurately.Then,when building the matrix factorization model,this thesis considers the data association among clusters,and synthesize the feature information of users and items from the local and global point of view,so that the isolated sub-matrix factorization model is related to each other.On this basis,this thesis proposes a matrix factorization recommendation algorithm based on the relationship between clusters.The experimental results show that the proposed algorithm based on clustering relationship can make full use of the local and global associations between users and items,and effectively improve the performance of the recommendation system.