IWDA: Importance Weighting for Drift Adaptation in Streaming Supervised Learning Problems.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Distribution drift, particularly concept drift in streaming machine learning (ML), degrades model performance when data distributions change over time.
• Existing methods struggle with efficient adaptation to these evolving data patterns in online, nonstationary settings.

Purpose of the Study:

• To introduce a novel, learner-agnostic algorithm for drift adaptation in supervised online learning.
• To enable efficient retraining of ML models upon detecting concept drift using importance weighting.

Main Methods:

• Developed Importance Weighting for Drift Adaptation (IWDA), an algorithm that incrementally estimates joint probability densities.
• Employs importance-weighted empirical risk minimization for retraining when drift is detected.
• Calculates importance weights using estimated densities for all observed samples, maximizing information utilization.

Main Results:

• IWDA demonstrates efficient retraining by leveraging all available data through importance weighting.
• Theoretical analysis supports the algorithm's effectiveness in abrupt drift scenarios.
• Numerical simulations show IWDA often outperforms state-of-the-art stream learning techniques on synthetic and real-world data.

Conclusions:

• IWDA offers an effective solution for adapting machine learning models to concept drift in streaming data.
• The algorithm's ability to efficiently utilize historical data makes it a competitive approach compared to adaptive ensemble methods.