Split Difference Weighting: An Enhanced Decision Tree Approach for Imbalanced Classification

Abstract

Imbalanced data classification remains a significant challenge in machine learning, particularly in decision tree algorithms where majority class features are often overshadowed. This study introduces a novel split index based on class key decision factor (CKD factor) to address this issue. We propose two new algorithms: Split Difference Decision Tree (SDDT) and Weighted Split Difference Classification and Regression Tree (WSD-CART). These algorithms enhance feature expression for majority classes during node splitting, thereby improving classification performance on imbalanced datasets. Experiments conducted on five UCI datasets with varying imbalance levels demonstrate the effectiveness of our approach. The WSD-CART algorithm consistently outperformed traditional methods, showing significant improvements in F-score, AUC, precision, recall, and accuracy, particularly for majority classes. In a real-world application to space product material classification, our method increased the true positive rate for majority class identification from 66.32% to 76.17%, while maintaining high overall accuracy. This study contributes to the field of imbalanced learning by providing a new perspective on decision tree split criteria. The proposed methods offer both improved classification performance and interpretable decision rules, making them valuable for various domains dealing with imbalanced data.