Collaborative Filtering Recommendation Algorithm Based On Kernel Matrix Factorization And Robust Estimation

Collaborative filtering recommendation algorithm is a recommendation technology in recommender systems, but in the face of shilling attacks, there are some problems for collaborative filtering algorithms. Firstly, the filtering recommendation algorithm based on traditional matrix factorization, which has weaker tolerance of outlier, and user feature matrix and item feature matrix are influenced by attack profiles. So the model’s anti-attack is lower. Then, most algorithms have weaker anti-sparse and the robustness. In order to guarantee that the quality of recommendation is well, the paper has studied robust model of algorithms on the basic of the existing research, aiming at enhancing recommendation accuracy and robustness. The main contributions of this paper are as follows:Firstly, in the paper, we propose a novel robust recommendation algorithm based on kernel matrix factorization. Then, we construct a robust kernel matrix factorization model for collaborative recommendation by using kernel mapping of the rating matrix and kernel distance and regulate residual error with the scale factor, which can enhance the power of the model’s anti-attack and realize the robust estimation of user feature matrix and item feature matrix. Finally, we introduce kernel distance to compute the similarity between users in order to improve the credibility of user similarity and reduce the influence of attack profiles on the recommendation results.Secondly, in the paper, we propose a novel robust recommendation algorithm based on kernel regularization and weighted M-estimator. In loss function, we introduce Gaussian kernel function as M-estimator function that realize the robust parameters estimation,as well as applying kernel function to similarity computation to handle the nonlinearity among similar users, importantly, we also introduce related knowledge of Correntropy Induced Metric(CIM), which has better robustness, and we serve user feature matrix and item feature matrix.that are handled with CIM as recommendation, so we strengthen anti-sparse of the algorithm and reduce the influence of attack profiles on user feature matrix and item feature matrix.Finally, we conduct experiments on the Movie Lens dataset to demonstrate the effectiveness of the proposed algorithm comparing to other algorithms.