Adaptive Chunk-Based Dynamic Weighted Majority for Imbalanced Data Streams With Concept Drift.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Concept drift significantly impacts classification stability in streaming data, especially with imbalanced datasets.
• Existing ensemble methods struggle to balance stability and adaptability, and fixed-size chunks exacerbate issues with imbalanced data.

Purpose of the Study:

• To propose an adaptive chunk-based incremental learning method (ACDWM) for imbalanced streaming data with concept drift.
• To enhance the stability and adaptability of online learners in dynamic environments.

Main Methods:

• ACDWM employs an ensemble framework with dynamically weighted classifiers based on their performance on current data chunks.
• Chunk size is adaptively determined using statistical hypothesis tests to ensure classifier stability.
• The method is entirely incremental, requiring no storage of previous data.

Main Results:

• ACDWM demonstrates superior performance over state-of-the-art chunk-based and online methods on synthetic and real-world datasets.
• The method effectively maintains stability during non-drifted streams and rapidly adapts to new concepts.
• Experiments confirm ACDWM's efficiency due to storing a limited number of classifiers.

Conclusions:

• ACDWM offers a robust solution for handling concept drift and imbalanced data in online learning scenarios.
• The adaptive chunk sizing and dynamic weighting contribute to improved classification accuracy and adaptability.
• The proposed method provides an efficient and incremental approach for real-time data stream analysis.