Adaptation Strategies for Automated Machine Learning on Evolving Data.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Automated Machine Learning (AutoML) systems excel at model creation for static datasets.
• The adaptability of AutoML to evolving data streams and concept drift remains a significant challenge.
• Understanding and mitigating concept drift is crucial for maintaining model performance over time.

Purpose of the Study:

• To investigate the impact of concept drift on AutoML performance.
• To identify and evaluate effective strategies for adapting AutoML to data changes.
• To propose enhancements for robust AutoML techniques in dynamic environments.

Main Methods:

• Proposed six novel concept drift adaptation strategies.
• Evaluated strategies across diverse AutoML approaches: Bayesian optimization, genetic programming, and random search with automated stacking.
• Conducted empirical evaluations on both real-world and synthetic data streams exhibiting various concept drift types.

Main Results:

• Demonstrated that specific adaptation strategies significantly improve AutoML robustness against concept drift.
• Identified variations in strategy effectiveness depending on the type of concept drift and the AutoML method used.
• Quantified performance degradation of standard AutoML without adaptation strategies.

Conclusions:

• Concept drift poses a considerable challenge to AutoML systems.
• The proposed adaptation strategies offer viable solutions for enhancing AutoML robustness.
• Further research into adaptive AutoML is essential for real-world applications with evolving data.