CHDS: a boundary-oriented oversampling method for imbalanced data based on convex hull and Delaunay triangulation.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Class imbalance is a common problem in machine learning.
• It causes biased models that perform poorly on minority classes.
• Existing methods struggle to effectively address this challenge.

Purpose of the Study:

• To introduce a novel oversampling framework called Convex Hull Delaunay Sampling (CHDS).
• To improve minority class classification accuracy in imbalanced datasets.
• To provide a reliable data-level solution for risk-sensitive applications.

Main Methods:

• CHDS identifies boundary minority samples using local neighborhood entropy.
• It constructs convex hulls and uses Delaunay Triangulation for safe region selection.
• Synthetic samples are generated within these regions using a Dirichlet distribution strategy.

Main Results:

• CHDS significantly outperformed 15 other resampling methods across 28 imbalanced datasets.
• Performance improvements were observed in AUC, F1-score, G-mean, and Recall.
• Ablation studies confirmed the effectiveness of each component of the CHDS framework.

Conclusions:

• CHDS offers a robust data-level solution for class imbalance problems.
• The framework enhances decision-making reliability in critical applications.
• It effectively models the minority class distribution for improved synthetic data generation.