Research On Recommendation Technology Based On Matrix Factorization Model

The widespread use of personalized recommendations greatly improve the user's ability to interact with the website.Not only do websites recommend effective information to users accurately,but also users find valuable information on their own in a shorter period of time.Therefore,it has aroused wide public concern and exploration of business community and academic society.Model-based recommendation technology is one of the most widely used technologies in the recommended system,which is also a hotspot in the field of recommendation.This method has the advantages of accurate prediction,good scalability,and good adaptability in a variety of real-world scenarios.In this paper,the collaborative filtering recommendation technique based on factorization is studied in detail,and the main achievements are as follows:(1)Based on the traditional recommendation method which is to apply factorization on the user-item scoring matrix,an improved model for the low rank approximation of the adjacency matrix of the relational graph for Top-N recommendation is proposed.Starting from the consistency of the similarity space between the user and the item,this method constrains two feature matrices that it is possible to avoid local over-fitting to a certain extent while preserving the consistency relation between the high-dimensional representation space and the low-dimensional recessive space.The performance of collaborative filtering recommendation system based on relational graph is improved effectively;(2)The traditional method of label information as bias to guide the user preference model,only a small range of adjustment score matrix,this paper further tap the label information,recommending the decomposition method of factor fusion tag information,the label information is divided into user label information and project tag information,and mapped into a low dimensional space as constraints on the user characteristic matrix and characteristic matrix of project,not only can increase the diversity of data,alleviate the calculation problem of high dimension and sparse data brings,and make more personalized recommendation results;(3)The recommended method of strengthening user portrait of discrimination,the recommended method of decomposing cluster information into the matrix factor,users and user portrait feature matrix of hidden constraints,strengthen the discrimination of user portrait,produce correction effect,make it accord with the characteristics of user preferences,do more practical application scenarios,the experimental results have been proved.