Regularized singular value decomposition in news recommendation system

Abstract

Singular Value Decomposition was widely used in recommendation system because of the Netflix Prize competition. The method decomposed the user item rating matrix into two matrix with low rank. In order to avoid overfitting the observed user item ratings. It used l2 regularization method to regularize the learned parameters by penalizing their magnitudes. It can solve the problem of sparsity and reduce the dimension of user item rating matrix. It obtain good result using the Root Mean Square Error (RMSE) as evaluation index. But the method cost a lot of time. In this paper, we proposed l1 regularization method and combine l1 and l2 regularization method to regularize the learned parameters of SVD. l1 regularization method show great superiority in the problem of sparsity. Experimental results on XMU News data set and Movie lens data set demonstrate the efficiency and effectiveness of the proposed model. l1 regularization method can represent the users' and items' implicit relation with fewer feature. Combining l1 and l2 regularization method perform well on the RMSE and costing time.