Research On Hybrid Recommendation Algorithm Based On Matrix Factorization And Feature Analysis

With the rapid development of the mobile Internet,a wide variety of goods are available on various e-commerce platforms.In the face of many kinds of goods,it is difficult for users to find the goods that meet their needs.Therefore,the development of high-quality personalized commodity recommendation system has become the only way to the development of mobile e-commerce platform.However,the existing traditional collaborative filtering algorithm can't support now huge amount of data,and collaborative filtering algorithms due to the limitation of the proposed times,not fully consider the characteristics of the user and the mobile electronic commerce,and recommend low accuracy and influenced by sparse sex big and poor stability problem,recommend the results,the accuracy and efficiency of impact the user's shopping experience.Therefore,combining with the emerging technologies in the era of big data,it has become an urgent problem for the academia and enterprises to improve the traditional recommendation algorithm and make it meet the current recommendation needs of mass data.Based on the characteristics of users' shopping behavior and the characteristics of mobile e-commerce,this paper presents a hybrid recommendation algorithm based on matrix decomposition and feature analysis.The algorithm is divided into two phases,recall and sorting,first in the recall of data preprocessing,converts a variety of interactive behavior of implicit feedback to intuitive quantifiable explicit rating,and then build the user and the score matrix,through joint clustering,using the theory of graph module maximum of sparse matrix can be divided into a number of low rank evaluation of molecular matrix,and using matrix decomposition to deal with each of the low rank evaluation matrix molecules fill the predicted value.In the matrix decomposition stage,the improved non-negative matrix decomposition is adopted,the eigenvalue selection ability is improved by introducing the L1 norm,and the model overfitting is prevented by the L2 norm.Recall layer algorithm is designed to use less features quick recall the candidate set,in the candidate with more features to build precise recommendation model,so the sorting characteristics of engineering is utilized to extract the user layer and commodity characteristic vector,using the factorization machine machine learning methods and depth of the neural network to build the potential relationship between the characteristics and the score,the factorization machine low-order feature extracting,the depth of the neural network to extract the higher-order features combination.Finally,the sorting function is completed in the candidate set to produce recommendation results.This model integrates matrix decomposition and factorization machine deep learning,makes use of the characteristics of simple and fast matrix decomposition,and also fully considers the accurate recommendation based on deep learning,which is helpful to improve the recommendation effect.In this paper,the data set of Jdata algorithm competition is used to design experiments to verify the accuracy,coverage and stability of the proposed algorithm,and the recommended quality is compared with the traditional baseline algorithm.Experimental results show that the proposed algorithm has a great improvement in recommendation quality and efficiency compared with the baseline algorithm,and the algorithm supports the processing of massive data,which proves the effectiveness of the proposed algorithm.