MEXT: A Parameter-Free Oversampling Approach for Multi-Class Imbalanced Datasets

Machine learning classifiers face significant challenges when confronted with class-imbalanced datasets, particularly in multi-class scenarios. The inherent skewness in class distributions often leads to biased model predictions, with classifiers struggling to accurately identify instances from underrepresented classes. This paper introduces MEXT, a novel parameter-free oversampling technique specifically designed for multi-class imbalanced datasets. Unlike conventional approaches that often rely on the one-against-all strategy and require manual parameter tuning for each class, MEXT addresses these limitations by simultaneously balancing all classes. By leveraging anomalous score analysis, MEXT automatically determines optimal locations for synthesizing new instances of minority classes, eliminating the need for manual parameter selection. The technique aims to achieve a balanced class distribution where each class has an equal number of instances. To evaluate MEXT’s effectiveness, the experiments were conducted extensively on a collection of multi-class datasets from the UCI repository. The proposed MEXT algorithm was evaluated against a suite of state-of-the-art SMOTE-based oversampling techniques, including SMOTE, ADASYN, Safe-Level SMOTE, MDO, and DSRBF. All comparative algorithms were implemented within the one-against-all framework. Hyperparameter optimization for each algorithm was performed using grid search. An automated machine learning pipeline was employed to identify the optimal classifier-hyperparameter combination for each dataset and oversampling technique. The Wilcoxon signed-rank test was subsequently utilized to statistically assess the performance of MEXT relative to the other oversampling techniques. The results demonstrate that MEXT consistently outperforms the other methods in terms of average ranking of key evaluation metrics, including macro-precision, macro-recall, F1-measure, and G-mean, indicating its superior ability to address multi-class imbalanced learning problems.