A Novel Ensemble Method for Imbalanced Data Learning: Bagging of Extrapolation-SMOTE SVM

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Class imbalance is a common issue in real-world datasets, leading to suboptimal model performance.
• Existing methods like sampling and cost-sensitive learning have limitations in addressing this challenge.
• Samples near the decision boundary hold crucial discriminative information for improving model accuracy.

Purpose of the Study:

• To develop a novel synthetic minority oversampling technique that incorporates borderline information.
• To propose a new ensemble method, Bagging of Extrapolation Borderline-SMOTE SVM (BEBS), for imbalanced data learning (IDL).
• To address the limitations of existing methods by valuing and utilizing borderline samples.

Main Methods:

• Designed a new synthetic minority oversampling technique inspired by geometric principles to incorporate borderline information.
• Developed an ensemble model, Bagging of Extrapolation Borderline-SMOTE SVM (BEBS), combining ensemble of Support Vector Machines (SVMs) with extrapolation and borderline information.
• Utilized Bagging to enhance the robustness and complexity of the decision boundary.

Main Results:

• BEBS demonstrated significant superior performance on open-access imbalanced datasets.
• The proposed method effectively corrects the skew of the decision boundary by constructing synthetic samples near it.
• Experimental results validated the effectiveness of incorporating borderline information within an ensemble framework.

Conclusions:

• BEBS is a novel and effective approach for tackling imbalanced data learning problems.
• The study highlights the importance of borderline information and ensemble methods for improving model performance.
• This work represents the first model to combine ensemble of SVMs with borderline information for imbalanced data.