QuanDA: Quantile-Based Discriminant Analysis for High-Dimensional Imbalanced Classification.

Area of Science:

• Machine Learning
• Statistical Modeling
• Data Science

Background:

• Binary classification with imbalanced classes presents significant challenges for standard machine learning algorithms.
• High-dimensional data with low sample sizes exacerbates these issues due to noisy features and underrepresentation of the minority class.

Purpose of the Study:

• To introduce Quantile-based Discriminant Analysis (QuanDA) as a novel approach for imbalanced classification.
• To address the difficulties posed by low sample sizes and high dimensionality in classification tasks.

Main Methods:

• Developed Quantile-based Discriminant Analysis (QuanDA), leveraging a novel connection with quantile regression.
• Utilized appropriately chosen quantile levels to naturally account for class imbalance.
• Provided theoretical analysis for validation in ultra-high dimensional settings.

Main Results:

• QuanDA demonstrates superior performance compared to existing imbalanced classification methods.
• Outperformed cost-sensitive large-margin classifiers, random forests, and SMOTE in simulations and benchmark data analysis.
• Effectively handles class imbalance and high-dimensional data with low sample sizes.

Conclusions:

• Quantile-based Discriminant Analysis (QuanDA) offers a robust solution for imbalanced classification problems.
• The method shows significant promise for improving predictive performance in challenging data settings.
• QuanDA provides a theoretically validated and empirically superior alternative for imbalanced data analysis.