Driftage: a multi-agent system framework for concept drift detection.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Machine learning algorithms degrade in accuracy due to rapid societal data and behavior changes.
• This degradation, known as concept drift, necessitates complex and costly detection and maintenance strategies.
• Existing methods often require specialized knowledge in drift detection algorithms and software engineering.

Purpose of the Study:

• To introduce Driftage, a novel framework designed to simplify the implementation of concept drift detectors.
• To enhance the explainability of concept drift detection by dividing responsibilities among agents.
• To enable more dynamic adaptation of machine learning algorithms to evolving data patterns.

Main Methods:

• Development of Driftage, a new framework utilizing multi-agent systems.
• Implementation of a case study using electromyography muscle activity monitoring.
• Demonstration of agent-based responsibility division for concept drift detection.

Main Results:

• Significant simplification in implementing concept drift detection.
• Enhanced interpretability of drift detection processes.
• Reduction in false-positive drift detections.
• Improved interactivity of drift detectors with external knowledge bases.

Conclusions:

• Driftage establishes a new paradigm for implementing concept drift algorithms.
• The multi-agent architecture facilitates distributed drift detection responsibilities.
• This approach leads to more interpretable and dynamically adaptable algorithms.