Evaluating the Impact of Data Balancing Techniques on the k-Nearest Neighbors Algorithm for Microarray Data Classification

Microarray data classification poses significant challenges in bioinformatics due to the nature of the data, which has a very high number of features but a limited number of samples, and an unbalanced class distribution. This condition can cause a decrease in the performance of classification models, including k-Nearest Neighbor (kNN). This study aims to evaluate the performance of the kNN algorithm in classifying unbalanced and balanced data. The balancing techniques used are Random Undersampling (RUS), Random Oversampling (ROS), and Synthetic Minority Over-sampling Technique (SMOTE). The datasets used in this study are three leukemia datasets with different class structures, namely two, three, and four classes. The experimental results show that the ROS and SMOTE techniques consistently improve the performance of kNN, with the best accuracy reaching more than 97%. In the two-class dataset, ROS gave the best performance (99.4%), while in the three-class dataset, SMOTE showed the most optimal results (98.5%). In the four-class dataset, the performance improvement due to balancing was very significant; SMOTE and ROS were able to improve the accuracy from 89.7% (without balancing) to 99.0% and 98.8%, respectively. Although RUS recorded perfect accuracy of 100%, the results were anomalous and inconsistent. RUS showed less stable performance and was often lower than the condition without balancing, especially on datasets with four classes. Overall, the SMOTE technique proved to be the most stable and effective for various class structures. This study shows the importance of balancing strategies in the classification of complex and imbalanced microarray data.