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..

Optimal Regret of Bandits under Differential Privacy

Authors: Achraf Azize, Yulian Wu, Junya Honda, Francesco Orabona, Shinji Ito, Debabrota Basu

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, our numerical experiments validate that DP-KLUCB and DP-IMED achieve lower regret than the existing ϵ-global DP bandit algorithms.
Researcher Affiliation Academia Achraf Azize Fair Play Joint Team CREST, ENSAE Paris EMAIL Yulian Wu King Abdullah University of Science & Technology (KAUST) Thuwal 23955-6900, Kingdom of Saudi Arabia EMAIL Junya Honda Kyoto University RIKEN AIP EMAIL Francesco Orabona King Abdullah University for Science & Technology (KAUST) Thuwal 23955-6900, Kingdom of Saudi Arabia EMAIL Shinji Ito The University of Tokyo RIKEN AIP EMAIL Debabrota Basu Univ. Lille, Inria, CNRS Centrale Lille, UMR 9189-CRISt AL EMAIL
Pseudocode Yes Algorithm 1: DP-KLUCB and DP-IMED
Open Source Code Yes Also, the full code to reproduce our figures is provided in the supplementary material.
Open Datasets No As our algorithm is straightforward to code, and the datasets consist of simulated Bernoulli instances, all the experiments could be reproduced using a commercial laptop. Our algorithms are implemented from scratch and are tested on synthetic Bernoulli data.
Dataset Splits No As our algorithm is straightforward to code, and the datasets consist of simulated Bernoulli instances, all the experiments could be reproduced using a commercial laptop.
Hardware Specification Yes We implement all the algorithms in Python (version 3.8) and on an 8 core 64-bits Intel i5@1.6 GHz CPU.
Software Dependencies Yes We implement all the algorithms in Python (version 3.8) and on an 8 core 64-bits Intel i5@1.6 GHz CPU.
Experiment Setup Yes As in Sajed and Sheffet [2019], Azize and Basu [2022], Hu and Hegde [2022], we consider 4 different 5-arm Bernoulli environments, with specific arm-means choices. We run each algorithm 100 times for T = 10^6. For ϵ = 0.25, we plot the mean regret in Figure 1 for µ1 [0.75, 0.7, 0.7, 0.7, 0.7] in the left and µ2 [0.75, 0.625, 0.5, 0.375, 0.25] in the right. In Appendix G, we present additional results for some other environments under different budgets. ...we chose n0 = 1 and α = 2 for DP-KLUCB and DP-IMED. Also, to comply with the regret analysis in [Azize and Basu, 2022, Sajed and Sheffet, 2019], we chose α = 3.1 in Ada P-KLUCB, and β = 1/T in DP-SE.