Practical Continual Mean Estimation via Matrix Factorization under User-Level Differential Privacy

For the problem of continual mean estimation under user-level differential privacy, the researchers introduce the Matrix Factorization mechanism into the approximate differential privacy setting and propose a dedicated factorization for mean estimation. Prior work on this problem focused on pure differential privacy, and the estimation noise was too large; the authors state that the new factorization is efficient and accurate, achieving an asymptotically lower mean squared error bound.