An Improved Version of Edited Nearest Neighbor Undersampling Method Based on the kNN Approach

In the field of machine learning, handling imbalanced datasets remains a critical challenge, often addressed through various sampling techniques. Among these techniques, the Edited Nearest Neighbor (ENN) undersampling method is widely recognized for its ability to enhance classifier performance by reducing class imbalance. However, the traditional ENN method has limitations, such as the removal of potentially informative instances and suboptimal performance in complex datasets. This paper presents an improved version of the ENN undersampling method, leveraging the k-Nearest Neighbors (kNN) approach to refine the selection process for instance removal. The proposed method improves upon the traditional ENN by incorporating a more sophisticated neighbor evaluation criterion based on the k-NN algorithm, which better preserves informative instances while effectively reducing noise. Through extensive experiments on multiple benchmark datasets, we demonstrate that our improved ENN method achieves superior performance in terms of classification accuracy, F1-score, and AUC, compared to the traditional ENN and other state-of-the-art undersampling techniques. The results indicate that the improved ENN method not only mitigates the class imbalance problem more effectively but also maintains a higher level of data integrity, thereby enhancingthe robustness and reliability of machine learning models. This advancement provides a valuable tool for practitioners dealing with imbalanced datasets, contributing to the development of more accurate and efficient predictive models.