Optimization Class Imbalance Using Undersampling Tomek Links on Backpropagation

Abstract

lass imbalance is a prevalent issue in machine learning, where the majority class significantly outweighs the minority class, leading to biased models that fail to generalize effectively. This imbalance can result in poor predictive performance, particularly for the minority class, which may contain critical information. To address this issue, this study applies Tomek Links, an undersampling technique, to balance the dataset before training Backpropagation Neural Networks for classification tasks. Undersampling is preferred over oversampling to reduce the risk of overfitting and improve model generalization by eliminating noise. The results indicate that applying Tomek Links significantly improves model performance. The Mean Square Error (MSE) decreases from 0.1686 to 0.1466, and accuracy increases from 74.68% to 80.42%. Additionally, the confusion matrix analysis shows an increase in True Positives (TP) from 23 to 34 and a decrease in False Negatives (FN) from 19 to 14, demonstrating better classification of the minority class. Furthermore, recall improves from 0.6545 to 0.7083, and the F1-score rises from 0.6605 to 0.6868. These findings highlight the effectiveness of Tomek Links as an undersampling method in improving neural network performance for class-imbalanced datasets.