Dimension Reduction Based on Effects of Experienced Users in Recommender Systems

摘要

The paradox of huge volume with high sparsity of rating data in collaborative filtering (CF) system motivates the present paper to utilize information underlying sparsity to reduce the dimensionality of data. This difference in user experiences resembles factor underlying widely used term frequency weighting scheme in information retrieval. Hypothesis of Rational Authorities Bias (H-RAB) is proposed, supposing that higher prediction accuracy can be attained to emphasize referential users with higher experiences. Dimension reduction suggests pruning all referential users with less experience than a given maturity threshold. Empirical results from a series of experiments on three major public available CF datasets justify the soundness of both modifications and validity of H-RAB. A few open issues are also proposed for future efforts.