Improved Algorithms Of Collaborative Filtering Based On Matrix Factorization

With the development of information technology,internet technology is also changing,enables us to contact with the vast amount of information,meeting the needs for information.At the same time,the rapid development of the Internet also makes the information explosive growth,leading to the emergence of information overload.Search engine is the way to ease the problem of information overload.However,it is useless when the user can not provide a clear key words;at the same time,for the same keywords,the search results are the same,can not meet user personalizational needs.As the representative of information filtering technology,the recommended system is a good solution to these problems.The research of recommendation system began in 1990s,a large number of research results have been appeared,and many kinds of recommendation algorithms have improved the accuracy of recommendation.In this paper,we mainly research the recommendation algorithm of collaborative filtering based on matrix factorization,based on the analysis of the classical matrix factorization collaborative filtering algorithm,an improved strategy is proposed,solving the problems of the present methods,improving the accuracy.The main works in this dissertation are as follows:(1).The research background and the research progress of the domestic and foreign has been systematically elaborated.Moreover,the structure,the classification and the problems of the present methods,have been analyzed and discussed in detail.Furthermore,especially makes a detailed interpretation of the collaborative filtering algorithm,introduce the memory-based and model-based collaborative filtering algorithm.(2).Regularization technique is needed for the collaborative filtering recommendation algorithm based on the gradient descent matrix factorization to restrict the problems.Regularization parameter used in the loss function can control the trade-off between prediction accuracy of the model and overfitting avoidance to the training.A method of multiple regularization parameters is proposed.It can obtain the regularization parameters according to the user activity and item popularity,avoid the overfitting in the users and items with different numbers,and get better prediction accuracy.The experiment results show that this method is correct and feasible.(3).The matrix factorization models which profile both users and items latent factors directly,and the neighborhood models which analyze similarities between users and items,are current research focuses of collaborative filtering.To solve the lack of these two models,a method of merging both matrix factorization models and neighborhood models is proposed,which can make further accuracy improvements.The experiment results show that this method is correct and feasible.