Enhancing privacy while preserving the accuracy of collaborative filtering

Collaborative Filtering (CF) is considered a powerful technique for generating personalized recommendations. Centralized storage of user profiles in CF systems presents a privacy breach, since the profiles are available to other users. Recent works proposed enhancing the privacy of the CF by distributing the profiles between multiple repositories. This work investigates how a
decentralized distributed storage of user profiles combined with data perturbation techniques mitigates the privacy issues. Experiments, conducted on three datasets, show that relatively large parts of the profiles can be perturbed without hampering the accuracy of the CF. The experiments also allow conclusion to be drawn regarding the specific users and parts of the profiles that are valuable for generating accurate CF predictions.