Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Online robust locally differentially private learning for nonparametric regression
Authors: Chenfei Gu, Qiangqiang Zhang, Ting Li, Jinhan Xie, Niansheng Tang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments validate our theoretical findings, demonstrating that our methods achieve strong robustness and privacy protection without sacrificing efficiency. In this section, we propose the online robust nonparametric estimation within the RKHS framework to the minimization problem (1), which is universal for non-privacy-preserving and privacy-preserving settings. This section evaluates the finite-sample performance of the proposed H-FSGD and PH-FSGD estimators under two cases: (Case 1) the true function f (x) is a sine function; (Case 2) f (x) is a linear combination of two Beta density functions. |
| Researcher Affiliation | Academia | 1School of Statistics and Data Science, Shanghai University of Finance and Economics 2Zhongtai Securities Institute for Financial Studies, Shandong University 3Yunnan Key Laboratory of Statistical Modeling and Data Analysis, Yunnan University EMAIL, EMAIL, EMAIL, EMAIL |
| Pseudocode | Yes | We summarize our approach as in Algorithm 1. We make the following remarks. Remark 1. Algorithm 1 supports mixed privacy regimes through iteration-specific privacy budgets. [...] We summarize the selection of Ο in Algorithm 2, and the proposed H-FSGD in Algorithm 3. Please see Appendix C. |
| Open Source Code | Yes | We provide open access to the code with sufficient instructions, as described in supplemental material. |
| Open Datasets | Yes | To better validate our method s practical utility, we conducted experiments on the real data Health and fitness dataset from Kaggle website 3. |
| Dataset Splits | Yes | We select 40000 samples as the training set and 1000 samples as the test set. |
| Hardware Specification | Yes | We compare the computational efficiency of different methods on a laptop equipped with a 3.20 GHz AMD Ryzen 7 5800H CPU and 16GB RAM. |
| Software Dependencies | No | Lastly, construct Ο = 1.345ΛΟ, where 1.345 is the default value in R package ( rlm function) to achieve 90% efficiency for normally distributed noise. |
| Experiment Setup | Yes | This section evaluates the finite-sample performance of the proposed H-FSGD and PH-FSGD estimators under two cases: (Case 1) the true function f (x) is a sine function; (Case 2) f (x) is a linear combination of two Beta density functions. Two types of errors are considered, the Gaussian distribution N(0, 0.25) and t distribution with degree of freedom 3. Simulation details are provided in Appendix F. We assess both non-private and privacy-preserving settings, repeating each setup 200 times independently and using the mean squared error (MSE) as the evaluation metric. Example 5.2 (Private synthetic data). In this example, we evaluate our PH-FSGD method under two LDP settings: (3, 0.1)-LDP and (2, 0.2)-LDP. We examine the H-FSGD performance under Case 1 for both constant and non-constant step size settings, with Ξ³0 [4, 24] and ΞΆ [0.3, 0.8]. |