New algorithm for recommender systems based on singular value decomposition method

Matrix factorization is one of the most favorable techniques based on model-based recommender systems. Matrix factorization approaches are superior than other algorithm of collaborative filtering for investigating sparsity data problem. In this paper, we develop a recommendation algorithm based on this idea that unknown ratings are affected from information which are extracted from available ratings, so, data need to preprocessing, since ratings of this dataset have categorical type, therefore, first impute suitable value to missing values for example replacing zero value of data with user median, item median, total median of ratings then, SVD approach is implemented on preprocessing data and predict rating of MovieLens dataset. New method is compared with simple SVD and normalize SVD methods on original data. Proposed methods are evaluated with three metrics: RMSE¹, RE², MAE³. We show that our work is efficient and significantly outperforms simple SVD and normalize SVD.