Accuracy analysis of recommendation system using singular value decomposition

Abstract

Recommendation systems use utility matrix to represent the user ratings for a particular items. But that matrix is sparse, that is, most of the user ratings are unknown. Predicting those unknown ratings is a big challenge of recommendation data mining task. Due to the sparse of data in utility matrix, few features become less important. Those features should be reduced to decline the computational complexity. Singular Value Decomposition (SVD) is a most powerful algorithm to predict unknown ratings by reducing the less significant features. Before applying SVD on utility matrix, all unknown ratings should be filled with some initial values. This paper focuses to generate two predictive matrixes by assigning two different initial values, where one is Zero and other is replacing unknown values with average item rating and then subtracting corresponding average user rating from the values. The accuracy of forecasted ratings has been justified over a sample dataset in this paper as well.