On the Performance of SVD-based Algorithms for Collaborative Filtering

In this paper, we describe and compare three Collaborative Filtering (CF) algorithms aiming at the low-rank approximation of the user-item ratings matrix. The algorithm implementations are based on three standard techniques for fitting a factor model to the data: Standard Singular Value Decomposition (sSVD), Principal Component Analysis (PCA) and Correspondence Analysis (CA). CA and PCA can be described as SVDs of appropriately transformed matrices, which is a key concept in this study. For each algorithm we implement two similar CF versions. The first one involves a direct rating prediction scheme based on the reduced user-item ratings matrix, while the second incorporates an additional neighborhood formation step. Next, we examine the impact of the aforementioned approaches on the quality of the generated predictions through a series of experiments. The experimental results showed that the approaches including the neighborhood formation step in most cases appear to be less accurate than the direct ones. Finally, CA-CF outperformed the SVD-CF and PCA-CF in terms of accuracy for small numbers of retained dimensions, but SVD-CF displayed the overall highest accuracy.