Robust Ensemble Learning Under Label Noise: A Theoretical Analysis and Framework-Specific Solutions

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Ensemble learning methods enhance decision models by combining weak learners.
• Real-world data often contains label noise, degrading supervised learning performance.
• Theoretical understanding of label noise impact on diverse ensemble frameworks is limited.

Purpose of the Study:

• Investigate the effects of label noise on ensemble learning frameworks.
• Analyze the theoretical underpinnings of performance degradation caused by noisy labels.
• Develop targeted strategies to mitigate label noise in ensemble methods.

Main Methods:

• Utilized bias-variance-diversity (BVD) decomposition theory.
• Examined the impact of noisy labels on Bagging, Boosting, and Stacking.
• Developed data filtering (DF), sample reweighting (SR), and interactive feature purification (IFPS) strategies.

Main Results:

• Characterized mechanisms of performance degradation due to label noise.
• Demonstrated consistent improvements with proposed strategies (DF, SR, IFPS) on benchmarks.
• Validated approaches on synthetic and real-world noisy-label datasets.

Conclusions:

• Label noise significantly impacts ensemble learning performance.
• Targeted strategies enhance ensemble robustness against noisy labels.
• Findings provide practical insights for applying ensemble methods to imperfect data.