Dual Geometry Margin Optimization for Coupled-Noisy Robust Ensemble Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Ensemble learning methods like Bagging and Boosting enhance model performance by combining base learners.
• Real-world datasets frequently contain feature and label noise, degrading model effectiveness, generalization, and leading to overfitting.
• Existing ensemble methods can be sensitive to noise, impacting their robustness and reliability.

Purpose of the Study:

• To propose a novel ensemble learning approach robust to both feature and label noise.
• To enhance the decision margin optimization inherent in ensemble methods for improved noise resistance.
• To introduce Dual Geometry Margin Boosting (DGMB) as a solution for noisy datasets.

Main Methods:

• Developed Dual Geometry Margin Boosting (DGMB), a novel ensemble learning technique.
• Incorporated Decision Plane Margin (DPM) to improve class separation.
• Utilized Hyper-Sphere Margin (HSM) to filter noisy samples during training.

Main Results:

• DGMB demonstrated significant resistance to both feature and label noise.
• Experimental results showed DGMB maintains strong performance on noise-contaminated datasets.
• DGMB outperformed other robust ensemble methods in noise resilience evaluations.

Conclusions:

• DGMB offers a robust solution for ensemble learning in the presence of feature and label noise.
• The proposed method effectively enhances class separation and filters noisy samples.
• DGMB represents a significant advancement in building reliable ensemble models for real-world, noisy data.