Immune centroids oversampling method for binary classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Imbalanced learning poses challenges for classification algorithms, where minority classes are often underrepresented.
• Existing oversampling techniques like Synthetic Minority Oversampling Technique (SMOTE) may not adequately capture data density or group characteristics.

Purpose of the Study:

• To introduce a novel oversampling method, Immune Centroids Oversampling Technique (ICOTE), for improving imbalanced classification.
• To enhance ICOTE by integrating it with Edited Nearest Neighbors (ENN) to form ICOTE + ENN, aiming for better class separation.

Main Methods:

• ICOTE utilizes an artificial immune network to generate synthetic examples (immune centroids) within high-density clusters of the minority class.
• The proposed ICOTE + ENN method combines ICOTE's synthetic data generation with ENN's noise removal of majority class instances encroaching on minority class regions.

Main Results:

• ICOTE effectively broadens decision regions for the minority class by generating representative synthetic examples.
• Experimental results demonstrate that both ICOTE and ICOTE + ENN outperform established resampling methods in classification tasks.
• The ICOTE + ENN combination shows particular efficacy in separating overlapping classes.

Conclusions:

• The proposed ICOTE method offers an effective approach to address data imbalance by generating targeted synthetic data.
• Integrating ENN with ICOTE further refines the dataset by removing noisy majority class examples, leading to superior classification performance.