Improving Rating Predictions by Baseline Estimation and Single Pass Low-Rank Approximation

Abstract

In this work, we propose a single pass low-rank matrix approximation technique for collaborative filtering. The unknown values in the partially filled ratings’ matrix is imputed by robust baseline prediction. The resulting matrix is not low-rank; but it is known from latent semantic analysis that the ratings matrix should be so since the number of factors guiding the users’ choice of items is limited. Following this analysis, we compute a low-rank approximation of the filled ratings matrix. This is a simple technique that requires computing the SVD only once—unlike more sophisticated matrix completion techniques. We compared our proposed method with state-of-the-art matrix completion and matrix factorization-based collaborative filtering approaches and found that our proposed method yields significantly better results. The mean absolute error (MAE) from competing techniques is around 78 % where as ours is around 74 %.