FADEL: Ensemble Learning Enhanced by Feature Augmentation and Discretization

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Data augmentation techniques like SMOTE and CTGAN are prevalent for imbalanced classification but can introduce bias, noise, and computational overhead.
• Existing methods may lead to overfitting, reduced predictive performance, and increased cybersecurity risks.

Purpose of the Study:

• To introduce FADEL, a novel architecture designed to overcome limitations of data augmentation in imbalanced classification.
• To improve minority class recognition and model stability without relying on data-level balancing or augmentation.

Main Methods:

• FADEL integrates feature-type awareness with a supervised discretization strategy.
• It employs a unique feature augmentation ensemble framework processing continuous and discretized features concurrently.
• The architecture dynamically routes feature sets to compatible base models.

Main Results:

• FADEL achieved 90.8% recall and 94.5% G-mean on an internal test set, without data augmentation.
• On an external validation set, FADEL maintained 91.9% recall and 86.7% G-mean.
• Results surpassed conventional ensemble methods trained on CTGAN-balanced datasets.

Conclusions:

• FADEL offers a robust solution for extreme class imbalance using feature augmentation, outperforming data augmentation approaches.
• The architecture demonstrates superior stability, computational efficiency, and cross-institutional generalizability.
• It provides a practical alternative to traditional data augmentation for imbalanced classification problems.