Abstract

We consider fingerprinting methods for collaborative filtering (CF) systems. In general, CF systems show their real strength when supplied with enormous data sets. Earlier work already suggests sketching techniques to handle massive amounts of information, but most prior analysis has so far been limited to non-ranking application scenarios and has focused mainly on a theoretical analysis. We demonstrate how to use fingerprinting methods to compute a family of rank correlation coefficients. Our methods allow identifying users who have similar rankings over a certain set of items, a problem that lies at the heart of CF applications. We show that our method allows approximating rank correlations with high accuracy and confidence. We examine the suggested methods empirically through a recommender system for the Netflix dataset, showing that the required fingerprint sizes are even smaller than the theoretical analysis suggests. We also explore the of use standard hash functions rather than min-wise independent hashes and the relation between the quality of the final recommendations and the fingerprint size.