UkM: A Novel Undersampling Method Using Modified k-Medoids Algorithm

Learning from imbalanced data remains a persistent challenge in classification tasks, often resulting in biased model performance and poor generalization, particularly for the minority class. This study proposes a novel undersampling technique, UkM, which integrates a modified k-medoids clustering algorithm to effectively balance class distributions. The goal of UkM is to balance an imbalanced dataset using a clustering-based approach and then classify the balanced dataset. Unlike conventional undersampling methods, UkM clusters the majority dataset and selectively reduces the majority class within each cluster to preserve data structure while improving class balance. The proposed method was evaluated on 32 publicly available imbalanced datasets and compared with seven established undersampling techniques—DBU, PUMD, NearMiss, RUS, UBMD, CLbU, and FCA—as well as an oversampling method, SMOTE. The datasets were balanced using the compared methods and analyzed using Random Forest and Multilayer Perceptron classifier with four evaluation metrics: F1-Score, Recall, G-mean, and AUC. Furthermore, a statistical analysis involving the Friedman test and the Nemenyi post-hoc test was conducted to determine the significance of performance differences. Computations were performed on two different dataset groups. The second group has a higher imbalance ratio. Based on the results of the first group, although SMOTE generally demonstrates the highest overall performance on average, UkM’s performance is noteworthy across multiple metrics and stands out as a prominent alternative, particularly among undersampling methods. UkM demonstrates high and consistent performance in terms of balanced classification success and overall accuracy and class discrimination based on F-score and AUC results, while exhibiting inconsistency and potential risk of failure in minority class recognition and maintaining class balance according to Recall and G-Mean metrics. The Nemenyi post-hoc analyses reveal that in most cases, there is no statistically significant difference between UkM and SMOTE, indicating that their performance can be considered comparable. Based on the results of the second group, while SMOTE shows the highest overall performance, UkM frequently outperforms other undersampling methods across several metrics. Although the Friedman test indicates significant differences in F-score among methods, pairwise Nemenyi tests mostly do not, suggesting comparable performance levels between the models.