Federated Bayesian optimization on random Fourier additive margin features and random kernel mapping

Abstract

Bayesian Optimization (BO) is an advanced technique for hyperparameter tuning in AutoML, particularly for optimizing black-box functions. This study mainly proposes the RAF kernel for Gaussian Processes and introduces two novel algorithms$:$ the Federated Bayesian additive marginal Thompson Sampling algorithm (FAT) and the Federated Bayesian random kernel Thompson Sampling algorithm (FAKT), the latter combining RAF with Random Fourier Features (RFF). To enhance privacy, we further develop DP-FAT and DP-FAKT by integrating Differential Privacy, which can reduce the communication costs while safeguarding client data. Experiments show that FAT and FAKT converge 10 communication rounds faster than existing methods (e.g., FTS), significantly improving efficiency in federated black-box optimization. These advancements demonstrate strong potential for large-scale learning tasks with enhanced privacy and reduced overhead.