Research and Publications

Introduction

This paper introduces Dynamic Ensemble Member Selection (DEMS), a framework that improves data stream classification by dynamically selecting the most effective subset of classifiers for each prediction. By ranking ensemble members based on estimated accuracy and predictive margin, DEMS adapts in real time to evolving data, enhancing predictive performance while keeping computational overhead low.

Problems addressed 

• Ensemble methods often treat all base learners equally, even when some become outdated under concept drift.

• Fixed‑size ensembles may include weak or redundant classifiers that reduce accuracy.

• Many streaming ensembles lack mechanisms to adaptively choose the best subset of models for each instance.

• Real‑time systems require improved accuracy without significant increases in runtime.

Methods overview

DEMS dynamically ranks all base learners in an ensemble using estimated accuracy and predictive margin, selecting only the top‑K classifiers for each prediction. The value of K is optimized automatically through a self‑adaptive mechanism that responds to changes in the data stream. This approach can be applied on top of existing ensemble algorithms, such as Streaming Random Patches, Adaptive Random Forest, and Online Smooth Boost, allowing them to adapt more effectively to concept drift while maintaining computational efficiency.

Key outcomes

• Introduces a dynamic, per‑instance ensemble selection strategy for streaming data.

• Improves accuracy across several leading ensemble algorithms.

• Maintains low computational overhead suitable for real‑time systems.

• Enhances adaptation to concept drift by prioritizing the most reliable classifiers.

• Demonstrates that selective ensemble activation can increase diversity and robustness.

Experimental findings

Experiments show that DEMS consistently improves classification accuracy across multiple state‑of‑the‑art streaming ensemble methods, even under challenging drift scenarios. Despite its dynamic selection process, DEMS introduces only a minimal runtime overhead of 11.66%, demonstrating that significant performance gains can be achieved without compromising real‑time processing requirements.