Research on an extended SVD recommendation algorithm based on user's neighbor model

The fact that Singular Value Decomposition(SVD) algorithm can be used to reduce dimension and remove noise with good scalability and accuracy makes it been widely adopted in recommendation systems. However, the valuable association information between user-user, item-item are neglected in traditional SVD. Therefore, we proposed a new extended SVD algorithm, named USVD in this paper to improve the algorithm's accuracy. USVD corrects SVD's prediction error of the target user according to the known prediction errors of his neighbors. Because SVD may lose user similarity information during the process of Matrix decomposition, we modified the least squares problem of SVD. Meanwhile, we also made improvement on the calculation of traditional Pearson Correlation Coefficient(PCCs). The proposed USVD is evaluated by extensive experiments with MovieLens datasets under RMSE index. Results show that USVD achieves better accuracy and higher efficiency compared to contemporary schemes.