Multi-Vendor Matrix Factorization with Differential Privacy

Recommender systems usually base their predictions on user-item interaction, a technique known as collaborative filtering. Vendors that utilize collaborative filtering generally exclusively use their own user-item interactions, but the accuracy of the recommendations may improve if several vendors share their data. Since user-item interaction data is typically privacy sensitive, sharing this data poses a privacy challenge for the collaborating vendors. In this work, we study the use of matrix factorization with multiple vendors under a differential privacy guarantee. Since differential privacy incurs a trade-off between privacy and utility, one obstacle is that the utility loss of the privacy-preserving measure may be greater than the utility gain of collaboration. We show that the empirical evaluation of this property in existing work is questionable, and that these works do not solve the problem. We also demonstrate that in a common experiment setup, the upper bound on the utility gain that can be achieved by collaboration is limited, which places a hard limit on the acceptable utility loss due to privacy preservation. This limit is small enough that even the utility loss in the current state of the art in differentially private matrix factorization in general exceeds it. We conclude with a number of open challenges for future work.Computer Science | Cyber Securit