Kernel-based dynamic ensemble approach for classifying imbalanced data with overlapping classes.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Imbalanced data and overlapping class boundaries pose significant challenges for traditional machine learning algorithms in real-world classification tasks.
• Existing methods often struggle to achieve high accuracy, precision, and G-mean when faced with these complex data distributions.

Purpose of the Study:

• To propose a novel Dynamic Ensemble Selection framework using a Boundary-Aware Kernel (DES-BAK) to enhance classification performance.
• To address limitations in handling imbalanced datasets and overlapping class boundaries.
• To improve key classification metrics including accuracy, precision, and G-mean.

Main Methods:

• Developed a Dynamic Ensemble Selection (DES) framework incorporating a Boundary-Aware Kernel (BAK).
• Introduced a novel boundary separation method within the kernel function to reduce class overlap.
• Utilized diverse feature representations and classification algorithms within the ensemble.
• Evaluated the framework on 15 benchmark datasets with imbalanced and overlapping constraints.

Main Results:

• The proposed DES-BAK framework demonstrated superior performance compared to several state-of-the-art methods.
• Significant improvements were observed in classification accuracy across various challenging datasets.
• The boundary separation method effectively reduced class overlap and enhanced ensemble capabilities.

Conclusions:

• The DES-BAK framework offers a robust solution for binary and multi-class classification problems with imbalanced data and overlapping boundaries.
• The integration of diverse classifiers and the novel boundary-aware kernel significantly boosts classification performance.
• This approach advances machine learning techniques for complex real-world applications.