A Distance-Based Weighted Undersampling Scheme for Support Vector Machines and its Application to Imbalanced...

Area of Science:

• Machine Learning
• Data Mining
• Pattern Recognition

Background:

• Support Vector Machines (SVM) are effective for classification and regression but struggle with imbalanced datasets and large sample sizes.
• Traditional undersampling methods for imbalanced data can lead to high computational costs and reduced accuracy due to random sampling and numerous iterations.

Purpose of the Study:

• To propose an improved Support Vector Machine algorithm, WU-SVM, that enhances classification performance on imbalanced datasets.
• To introduce a novel weighted undersampling (WU) scheme that leverages space geometry distance to better represent data distribution.

Main Methods:

• Developed a weighted undersampling (WU) scheme for SVM, assigning weights to majority class samples based on Euclidean distance to the hyperplane.
• Grouped majority samples into subregions (SRs) with higher-weighted samples prioritized in learning iterations to preserve data distribution information.
• Evaluated WU-SVM on 21 binary-class and 6 multiclass public datasets.

Main Results:

• WU-SVM demonstrated superior performance compared to state-of-the-art methods on imbalanced classification tasks.
• The algorithm showed significant improvements in key metrics including Area Under the Curve (AUC), F-Measure, and G-Mean.
• Experimental results confirmed the effectiveness of the proposed weighted undersampling approach in handling data imbalance.

Conclusions:

• The proposed WU-SVM algorithm effectively addresses the challenges of class imbalance and large-scale samples in Support Vector Machines.
• Weighted undersampling based on space geometry distance is a viable strategy for improving classification accuracy in imbalanced learning scenarios.
• WU-SVM offers a robust solution for imbalanced classification problems, outperforming existing techniques.