A Segment-Based Drift Adaptation Method for Data Streams.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Concept drift adaptation aims to update predictors for evolving data streams.
• Existing informed methods suffer from adaptation delay due to batch processing for drift detection.

Purpose of the Study:

• To propose a novel method for online concept drift adaptation that overcomes the adaptation delay.
• To introduce a new statistic, drift-gradient, for quantifying distributional discrepancy in real-time.

Main Methods:

• A sequentially updated statistic, drift-gradient, is proposed to quantify distributional discrepancy with each new instance.
• The Segment-Based Drift Adaptation (SEGA) method utilizes drift-gradient to online update predictors by retraining with segments exhibiting minimum drift.
• Drift-gradient is defined on training set segments, comparing old segments with the newest instance.

Main Results:

• SEGA effectively quantifies the increase in distributional discrepancy as new data arrives.
• The method enables online predictor updates by selecting segments with minimal drift-gradient.
• Extensive experiments demonstrate SEGA's superior performance over existing blind and informed drift adaptation techniques.

Conclusions:

• SEGA significantly reduces adaptation delay in concept drift scenarios.
• The drift-gradient statistic provides an effective measure for online drift detection and adaptation.
• SEGA offers a robust and efficient solution for adapting machine learning models to evolving data streams.