Improving the Variance of Differentially Private Randomized Experiments through Clustering

Authors: Adel Javanmard, Vahab Mirrokni, Jean Pouget-Abadie

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we evaluate the theoretical and empirical performance of our CLUSTER-DP algorithm on both real and simulated data, comparing it to common baselines, including two special cases of our algorithm: its unclustered version and a uniformprior version.
Researcher Affiliation Collaboration 1Marshall School of Business, University of Southern California, Los Angeles, USA 2Google Research, New York, USA. Correspondence to: Adel Javanmard <EMAIL>, Jean Pouget-Abadie <EMAIL>.
Pseudocode Yes Algorithm 1 UNIFORM-PRIOR-DP mechanism Algorithm 2 Our CLUSTER-DP mechanism
Open Source Code No The paper does not contain any explicit statement about providing source code, nor does it provide a link to a code repository.
Open Datasets Yes The You Tube social network dataset (Leskovec & Krevl, 2014) contains the friendship links of a set of users on You Tube, and the ground-truth clusters correspond to groups created by users.
Dataset Splits No The paper describes how data is generated and sampled for experiments (e.g., "super-population of three clusters of sizes 2.5e3, 5e3, and 10e4 units, and repeatedly draw uniformly at random sub-populations of three clusters from these original clusters"). For the You Tube dataset, it mentions "considering only the 50 largest communities". However, it does not provide specific training/test/validation dataset splits with percentages or counts, which are typically used for reproducing machine learning experiments.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory, or cloud instance types) used for running its experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions) needed to replicate the experiments.
Experiment Setup Yes Unless otherwise specified, and with no particular reason to fix parameters one way or another, we take K = 5, v = 5, and β = 4.5. We consider C = 3 clusters of sizes 500, 103, 2 103 with an equal number of controlled and treated units in each cluster. ... for CLUSTERDP mechanism, we set the truncation parameter γ = 0.02, the Laplace noise σ = 10, and the resampling probability λ = 0.8. ... For the CLUSTERDP and CLUSTER-FREE-DP, we set the Laplace parameter to σ = 10, and vary the truncation parameter γ [0.1/K, 1/K]. ... In the CLUSTER-DP mechanism, we set the truncation threshold to γ = 0.1/K and the Laplace noise level to σ = 5.