P3MCF: Practical Privacy-Preserving Multi-domain Collaborative Filtering

This paper proposes P3MCF, an efficient privacy-preserving, multi-domain collaborative filtering scheme for user oriented recommendations. P3MCF achieves a lightweight, high accuracy recommendation for a multi-domain recommendation system. In P3MCF, a data supplier transfers only statistical values on user ratings to recommenders in order to improve the accuracy of recommendations. P3MCF only requires transmission of O(m) statistical values for each data supplier, where m is the number of items in each user record. We implemented a prototype system and evaluated transaction time and accuracy of recommendations. Experiments confirmed that accuracy could be improved when using statistical values. The results also confirmed that the computation time for predicting a missing value was about 21 milliseconds if we use a public dataset where the number of ratings is 100,000. The experimental results demonstrated that P3MCF was sufficiently practical from the viewpoint of accuracy and transaction time. We also confirmed that P3MCF was applicable to several service models, such as a horizontally partitioned model and a vertically partitioned model.