Diverse Instance-Weighting Ensemble Based on Region Drift Disagreement for Concept Drift Adaptation

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Concept drift, changes in data distribution, is a challenge in evolving data streams.
• Ensemble learning with dynamic classifiers is effective for concept drift but maintaining diversity is difficult.

Purpose of the Study:

• To propose a new diversity measurement for ensemble learning in concept drift.
• To develop an instance-based ensemble learning algorithm to address concept drift.

Main Methods:

• A novel diversity measurement based on ensemble member agreement on regional distribution change probability.
• Instance weighting using estimations over regional distribution changes.
• Selection of disagreeing classifiers to maximize ensemble diversity.
• Development of the diverse instance-weighting ensemble (DiwE) algorithm.

Main Results:

• The proposed method effectively creates diversity by constructing different region sets.
• The DiwE algorithm demonstrates effectiveness and advantages in handling concept drift.
• Evaluations on synthetic and real-world data streams confirm the algorithm's performance.

Conclusions:

• The proposed diversity measurement and DiwE algorithm offer a robust solution for concept drift.
• This approach enhances ensemble learning performance in dynamic data stream environments.