Incremental learning of concept drift in nonstationary environments

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Concept drift in nonstationary environments (NSEs) poses challenges for machine learning models due to changing data distributions.
• Incremental learning algorithms are crucial for adapting to evolving data without retraining on historical data.

Purpose of the Study:

• To introduce Learn(++).NSE, an ensemble of classifiers-based approach for incremental learning in nonstationary environments.
• To develop an algorithm capable of handling various types of concept drift, including constant, variable, additive, and cyclical drift.

Main Methods:

• Learn(++).NSE employs an incremental learning strategy, processing data in batches without prior data access.
• It trains a new classifier for each data batch and combines them using dynamically weighted majority voting.
• The novelty lies in time-adjusted accuracy weighting for classifiers, enabling adaptation to current and past data distributions.

Main Results:

• The algorithm demonstrated effective tracking of changing data distributions across diverse simulated nonstationary environments.
• Learn(++).NSE showed robust performance regardless of the type or rate of concept drift encountered.
• Evaluations included synthetic datasets and a real-world weather prediction dataset, with comparisons to existing methods.

Conclusions:

• Learn(++).NSE provides a robust and adaptive solution for incremental learning in nonstationary environments.
• The dynamic weighting mechanism allows the algorithm to recognize and respond to evolving data patterns and reoccurring distributions.
• The study releases its datasets to facilitate further research and benchmarking in concept drift detection.