A Partition-Based Hybrid Algorithm for Effective Imbalanced Classification

Abstract

Imbalanced classification presents a significant challenge in real-world datasets, requiring innovative solutions to enhance performance. This study introduces a hybrid binary classification algorithm designed to effectively address this challenge. The algorithm identifies different data types, pairs them, and trains multiple models, which then vote on predictions using weighted strategies to ensure stable performance and minimize overfitting. Unlike some methods, it is designed to work consistently with both noisy and noise-free datasets, prioritizing overall stability rather than specific noise adjustments. The algorithm’s effectiveness is evaluated using Recall, G-Mean, and AUC, measuring its ability to detect the minority class while maintaining balance. The results reveal notable improvements in minority class detection, with Recall outperforming other methods in 16 out of 22 datasets, supported by paired t-tests. The algorithm also shows promising improvements in G-Mean and AUC, ranking first in 17 and 18 datasets, respectively. To further evaluate its performance, the study compares the proposed algorithm with previous methods using G-Mean. The comparison confirms that the proposed algorithm also exhibits strong performance, further highlighting its potential. These findings emphasize the algorithm’s versatility in handling diverse datasets and its ability to balance minority class detection with overall accuracy.