A Drift Region-Based Data Sample Filtering Method.

Area of Science:

• Machine Learning
• Data Mining
• Time Series Analysis

Background:

• Concept drift, changes in data distribution over time, renders machine learning models outdated.
• Detecting concept drift is crucial, but understanding the drift's location (drift understanding) is understudied.
• Effective drift understanding is necessary for adapting and updating models to new data patterns.

Purpose of the Study:

• To develop a novel method for drift understanding in data streams.
• To accurately identify the specific regions where concept drift occurs.
• To leverage drift region information for effective model updating and accurate tracking of new data patterns.

Main Methods:

• A drift region-based data sample filtering method is proposed.
• The method identifies drift regions to filter training data samples.
• Theoretical analysis guarantees uniform convergence of the identified drift region to the true drift region.

Main Results:

• The proposed method effectively identifies drift regions in data streams.
• Filtering data samples based on identified drift regions improves model performance.
• Experimental results show enhanced learning accuracy on datasets with concept drift.

Conclusions:

• The developed drift understanding method accurately identifies drift regions.
• The method provides a foundation for effective drift adaptation strategies.
• This approach significantly improves model accuracy in the presence of concept drift.