Abstract: To improve the accuracy of recommendation algorithm under differential privacy protection, a privacy preservation recommendation algorithm is proposed based on a clustering method with Bhattacharyya coefficient(BC). In the proposed algorithm, the Bhattacharyya coefficient is used as the standard of measuring item similarity. Based on the BC similarity, the items are clustered by K-medoids, and the private neighbors of the items are selected from the clusters. Finally, according to the selected nearest neighbor set, the user's rating is predicted and the Top-n recommendations are output. The proposed algorithm effectively overcomes the problem that the calculation of similarity must depend on the common rated ratings, improves the accuracy of the similarity measurement, and also avoid the problem of quality degradation of the nearest neighbor set due to privacy protection. It is shown that the proposed algorithm not only achieves privacy preservation but also guarantees the high quality of recommendation. Therefore, the proposed algorithm effectively balances the privacy preservation and the data utility, which has good application potential in the recommendation system.

Key words: collaborative filtering, Bhattacharyya coefficient, differential privacy preservation, K-medoids clustering, recommendation system