Improved Overlap-based Undersampling for Imbalanced Dataset Classification with Application to Epilepsy and...

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Imbalanced datasets are prevalent across various fields like healthcare and finance.
• Existing methods often focus on class distribution, neglecting the significant impact of class overlap.
• Class overlap poses a greater challenge to classification models than simple imbalance.

Purpose of the Study:

• To develop and evaluate novel overlap-based undersampling methods.
• To enhance the visibility of minority class instances within overlapping regions.
• To address the negative impact of class overlap on machine learning models.

Main Methods:

• Proposed overlap-based undersampling techniques utilizing soft clustering.
• Implemented an adaptable elimination threshold to identify and remove negative instances in overlap zones.
• Employed oversampling to emphasize minority class presence in borderline areas for improved clustering and detection.

Main Results:

• Demonstrated significant improvements in classification sensitivity.
• Achieved competitive performance compared to established and state-of-the-art methods.
• Validated effectiveness across simulated and real-world datasets with diverse imbalance and overlap levels.

Conclusions:

• Overlap-based undersampling is a promising approach for imbalanced classification.
• The proposed methods effectively handle class overlap, a critical factor in model performance.
• The techniques offer a valuable contribution to the field of imbalanced learning.