A0317
Title: Locally private nonparametric contextual multi-armed bandits
Authors: Yuheng Ma - Renmin University of China (China) [presenting]
Feiyu Jiang - Fudan University (China)
Zifeng Zhao - Notre Dame (United States)
Yi Yu - University of Warwick (United Kingdom)
Hanfang Yang - Renmin University of China (China)
Abstract: Motivated by privacy concerns in sequential decision-making on sensitive data, the challenge of nonparametric contextual multi-armed bandits (MAB) is addressed under local differential privacy (LDP). A uniform-confidence-bound-type estimator is developed, showing its minimax optimality supported by a matching minimax lower bound. The case where auxiliary datasets are available is further considered, subject also to LDP constraints. Under the covariate shift assumption on the auxiliary datasets, a jump-start scheme is proposed to effectively utilize the auxiliary data, and the minimax optimality is established by a matching lower bound. Comprehensive experiments on both synthetic and real-world datasets validate the theoretical results and underscore the effectiveness of the proposed methodology.
