Research and Publications

Introduction

This paper introduces Bayesian Stream Tuner, a hyperparameter optimisation method designed for real‑time, non‑stationary data streams. Unlike traditional tuning approaches that assume static data, BST adapts as the data distribution shifts, using online Bayesian modelling and drift detection to continually update hyperparameter choices.

Problems

• Existing HPO methods are offline, slow, and assume stationary data.

• Stream‑learning models need continuous, lightweight tuning.

• Concept drift makes fixed hyperparameters quickly suboptimal.

• Current online HPO methods lack global search and theoretical guarantees

Method

The study develops the Bayesian Stream Tuner (BST), an online hyperparameter optimisation method designed for evolving data streams. BST processes data in fixed windows, extracting lightweight statistical features that capture changes in the underlying distribution. These features are combined with hyperparameter descriptors and fed into an online Bayesian linear regression model, which continually updates its estimates of each configuration’s performance. The algorithm maintains a pool of candidate hyperparameter settings and periodically replaces weaker ones using a probability‑of‑improvement acquisition strategy. To remain responsive to concept drift, BST incorporates ADWIN drift detection, triggering resets of both the model and configuration pool when significant distribution shifts occur. The method is supported by theoretical regret bounds and evaluated across multiple streaming datasets for both classification and regression tasks.

Key outcomes

• Consistently outperforms state‑of‑the‑art online HPO methods across 20 datasets.

• Improves accuracy and reduces error for both classification and regression tasks.

• Adapts quickly after drift events.

• Delivers statistically significant gains despite modest computational overhead.

Findings

BST offers a principled, drift‑aware, globally informed approach to hyperparameter tuning in streaming environments. Its combination of Bayesian modelling, statistical features, and drift resets leads to more stable, accurate, and adaptive model performance than existing online tuning methods.