Document Type : Original Article

Authors

1 University of Mazandaran

2 University of mazandaran

Abstract

Support Vector Machine (SVM) is a powerful classification algorithm that separates samples by finding an optimal decision boundary. Its performance can degrade when feature variances differ across classes, potentially leading to suboptimal decision boundaries. A variance-weighted framework is proposed that reduces the influence of high-variance features while enhancing the impact of low-variance features, resulting in more accurate and robust decision boundaries. The method is applicable in both linear and nonlinear settings. Evaluation on synthetic datasets and real-world datasets, including Breast cancer and {\it a9a}, using cross-validation demonstrates that the variance-weighted SVM achieves higher accuracy and F1-score compared to soft SVM and LDM, particularly in scenarios with significant variance differences between classes.

Keywords