Algorithm Research Based On The Machine Learning And Matrix Factorization Of Recommender System

Matrix Factorization(MF)and its various approaches have been extensively researched in the field of recommendation systems over the past decade.Essentially,Matrix Factorization aims to find low-dimensional matrices that can most accurately represent the known rating scores.The two low-dimensional matrices generate by Matrix Factorization are always called the latent factors of user and the latent factors of item that would affect the rating of the user on the item.This thesis is based on the shortcomings of traditional recommendation system models in practical scenarios.This thesis combines matrix factorization and machine learning methods to study how to improve the prediction accuracy of recommendation system models.The main novelties and contents are follows:1.Traditional SVD models assume that all latent factors have the same weight,which may not be a realistic assumption in practical scenarios.This thesis proposed Double Weighted SVD(DWSVD)that each latent factor is associated with a corresponding weight parameter.This thesis compared the Double Weighted SVD model with other Matrix Factorization model on several public datasets based on the Root-Mean-Squared.The results show indicate the Double Weighted SVD performs better than Matrix Factorization model in all the experimental datasets.2.Traditional Matrix Factorization models assume that latent factors of user assign regularization weight λ_p and latent factors of item assign regularization weight λ_q.This approach may over-simplify the scenarios of recommendation systems.This Thesis proposed that model use functions instead of traditional regularization weight.Model probably should assign lower constraints to the latent factors associated with the items or users that reveal more information,and set higher constraints to the others.Based this methods,This thesis validates the proposed idea on different models and different public datasets.The experimental results show that the method proposed in this paper improves the accuracy of the traditional Matrix Factorization model.The method better predict rating information for long-tail items and alleviated the cold-start problem.3.Traditional Matrix Factorization models only model user and item information,without considering users’social network information.This thesis argues that users’implicit feedback information can be obtained from their social networks,and this information can be incorporated into latent vectors of user.This thesis uses sparse matrix to generate user social network relationships and compares the performance of different models on the dataset.The results show that social network information can improve the performance of recommendation systems.