Regression augmentation with data-driven segmentation.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Imbalanced regression occurs when target distributions are skewed, hindering model performance on underrepresented samples.
• Current methods often use arbitrary thresholds, failing to capture complex feature-target relationships for rare data.
• This limitation impacts various applications requiring accurate prediction of infrequent events.

Purpose of the Study:

• To develop a fully data-driven framework for imbalanced regression that automatically identifies and enriches minority samples.
• To overcome the limitations of fixed thresholds in existing data augmentation techniques.
• To improve the performance of machine learning models on skewed datasets.

Main Methods:

• A Generative Adversarial Network (GAN)-based augmentation framework is proposed.
• Mahalanobis-Gaussian Mixture Modeling (GMM) is utilized for automatic identification of minority samples.
• Deterministic nearest-neighbor matching is employed to enrich sparse data regions.

Main Results:

• The proposed method successfully identifies truly rare observations without relying on preset thresholds.
• Evaluated on 32 benchmark imbalanced regression datasets, the framework demonstrated superior performance.
• Outperformed existing state-of-the-art data augmentation techniques in imbalanced regression tasks.

Conclusions:

• The proposed GAN-based framework with Mahalanobis-GMM offers a robust, data-driven solution for imbalanced regression.
• This approach effectively addresses the challenge of underrepresented samples by accurately identifying and augmenting them.
• The method shows significant potential for improving machine learning model performance in real-world skewed data scenarios.