Reproduce “Differential Privacy in Kernelized Contextual Bandits via Random Projections” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Nikola Pavlovic, Sudeep Salgia, Qing Zhao
Researchers from Cornell and Carnegie Mellon Universities developed CAPRI, an algorithm for differentially private contextual kernelized bandits. The method provides cumulative regret guarantees for a wide range of kernels, including the Matérn family, under both Joint Differential Privacy (JDP) and Local Differential Privacy (LDP), aligning with optimal non-private regret bounds in the limit.
Key insights
- The methodology successfully extends provable differential privacy guarantees to a broad class of kernels, including the Matérn family, for cumulative regret in contextual bandits.
- The algorithm's regret bounds recover known order-optimal cumulative regret performance in the non-private limit and align with established lower/upper bounds for linear bandits.
- For stochastically generated contexts, the analysis suggests the absence of an O(T^(1/4)) cumulative regret gap between JDP and LDP, contrary to previous conjectures for adversarial contexts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.