Embedding Undersampling Rotation Forest for Imbalanced Problem

Area of Science:

• Machine Learning
• Data Science
• Computer Science

Background:

• Ensemble learning methods like Rotation Forest excel but struggle with imbalanced datasets.
• Imbalanced data, common in real-world scenarios, pose challenges due to skewed class distributions.
• Misclassifying minority class instances often carries higher costs than misclassifying majority class instances.

Purpose of the Study:

• To propose a novel ensemble learning method, Embedding Undersampling Rotation Forest (EURF), to effectively handle imbalanced datasets.
• To enhance the performance of Rotation Forest on datasets with a significant disparity between majority and minority classes.

Main Methods:

• EURF employs a two-step undersampling strategy within a feature space rotation framework.
• Majority class subsets are sampled to learn projection matrices, creating new feature spaces.
• Data is re-undersampled and projected into these new spaces to train individual classifiers, improving minority class recognition.

Main Results:

• Experimental results demonstrate that EURF significantly outperforms existing state-of-the-art methods on imbalanced datasets.
• The proposed undersampling techniques enhance the ability of individual classifiers to capture minority class features.
• EURF maintains classifier diversity while improving overall classification accuracy and reducing misclassification costs for the minority class.

Conclusions:

• EURF presents a robust and effective solution for ensemble learning on imbalanced data.
• The method successfully balances the need for classifier diversity with improved minority class performance.
• EURF offers a promising advancement in handling real-world imbalanced classification problems.