Research and Publications

Introduction

This paper introduces a regression framework for evolving data streams based on classical linear adaptive filtering, showing that simple recursive linear models can deliver fast, accurate predictions while naturally adapting to concept drift—without relying on ensembles or explicit drift detectors. By leveraging the inherent adaptability of linear filters, the authors demonstrate that these methods can outperform state‑of‑the‑art stream regression algorithms when the underlying problem is well‑approximated by a linear model.

Problems addressed

• Many streaming regression methods are adaptations of classifiers and often rely on complex ensembles to achieve accuracy.

• Existing approaches frequently require explicit drift detectors, adding overhead and sensitivity to parameter tuning.

• Nonlinear or ensemble‑based models can be computationally heavy, limiting their use in real‑time environments.

• There is a lack of simple, theoretically grounded regression methods that can naturally track drift in evolving data streams.

Methods overview

The authors revisit classical linear adaptive filtering theory and adapt its recursive formulations—such as LMS‑style updates—to the data stream regression setting. These recursive filters update model parameters incrementally and incorporate drift‑tracking mechanisms directly into the learning rule, allowing the model to adjust smoothly as the data distribution changes. This yields a lightweight, interpretable, and highly responsive regression method that avoids the need for ensembles or external drift detectors.

Key outcomes

• Introduces a simple, fast, and theoretically grounded approach to stream regression.

• Achieves strong performance without ensembles or explicit drift detectors.

• Naturally adapts to drift through recursive linear updates.

• Demonstrates competitive or superior accuracy on classical stream regression datasets.

• Highlights that linear models remain powerful when the underlying relationship is approximately linear.

Experimental findings

Across standard data stream regression benchmarks, the proposed linear adaptive filtering methods achieve competitive—and often superior—performance compared to state‑of‑the‑art streaming regressors, particularly in drifting scenarios where a linear model is appropriate. The results show that these simple recursive models can adapt quickly, maintain low computational cost, and outperform more complex methods under the right conditions.