A recommendation algorithm based on adaptive privacy protection

Differential privacy, as a provable privacy protection framework, has been widely applied in recommendation systems. However, the integration of existing collaborative filtering algorithms with differential privacy typically encounters two issues. Firstly, algorithms often use the same privacy budget for data from different time periods, introducing unnecessary noise and reducing performance. Secondly, when using gradient descent to solve collaborative filtering models, the allocation of privacy budgets is relatively simple, usually evenly distributed across each iteration. This allocation requires a predefined number of iterations. If the number is too small, the algorithm may terminate before reaching the optimal solution; if too large, each budget per iteration may be too small, leading to excessive noise. To address these issues, an adaptive privacy budget recommendation algorithm is proposed. Firstly, a dynamic time window-based privacy budget calculation algorithm is utilized to provide varying levels of privacy protection for rating data from different time periods, thus avoiding unnecessary noise. Then, an adaptive privacy budget gradient descent strategy is employed to solve the collaborative filtering model, which dynamically allocates privacy budgets based on the noise gradient of each round. This algorithm is compared with existing differential privacy-based recommendation algorithms on the Movielens-1M dataset, demonstrating superior predictive accuracy and lower sensitivity to privacy protection.