Boundary-Eliminated Pseudoinverse Linear Discriminant for Imbalanced Problems.

Area of Science:

• Machine Learning
• Data Science
• Computer Science

Background:

• Imbalanced data classification models are typically boundary-based or nonboundary-based.
• Existing methods face challenges in effectively handling skewed data distributions.

Purpose of the Study:

• To introduce a novel classification model that integrates advantages of both boundary-based and nonboundary-based approaches.
• To address limitations in existing models for imbalanced data classification.

Main Methods:

• A new boundary-eliminated (BE) model is proposed, partitioning the input space into three regions using two additional boundaries.
• The model employs a heuristic measurement for final decision-making, comparing test samples with selected training samples.
• The pseudoinverse linear discriminant (PILD) is integrated, creating the BEPILD classifier.

Main Results:

• The BEPILD classifier demonstrates effectiveness and efficiency in experiments.
• Performance was validated against 13 state-of-the-art methods.
• Evaluations were conducted on 31 imbalanced and 7 standard datasets.

Conclusions:

• The BEPILD model offers a promising advancement in imbalanced data classification.
• The novel approach of eliminating the original hyperplane and incorporating heuristic measurements proves effective.
• BEPILD shows competitive or superior performance compared to existing methods.