Accuracy of Privacy-Preserving Collaborative Filtering Based on Quasi-homomorphic Similarity

We study the problem of predicting a rating for an unseen item based on a distributed dataset owned by two honest-but-curious parties without revealing their private datasets to each other. Our proposed idea uses a new similarity measure such that the similarity aggregated from two local similarities is approximately equal to the global similarity. We evaluate the accuracy of prediction of rating and clarify the lower bound of estimation error and the expected value of error to be small enough to approximate the global prediction. We also show a new privacy preserving collaborative protocol with light weight overhead.