Improvement of Imbalanced Data Handling: A Hybrid Sampling Approach by using Adaptive Synthetic Sampling and Tomek links

Nowadays, the imbalanced data problem is still a crucial issue in classifying, especially in the medical data domain. For this reason, in this study, we propose a hybrid sampling method that combines adaptive synthetic sampling as an oversampling method and Tomek links as an undersampling method (ADASYN-Tomek) to improve the performance of the classification model. ADASYN has the advantage of synthesizing according to the hard-to-learn conditions of minor class samples. On the other hand, the Tomek links method can overcome class overlapping problems and remove noise. We conducted experiments on the pima indians diabetes data set obtained from the KEEL repository. The experimental results show that ADASYN-Tomek is better than the ADASYN or Tomek single method with a balanced accuracy score and F1score are 0.797 and 0.822, respectively. Furthermore, to improve the performance of the classification model, we performed tuning of the n-neighbors parameter in the oversampling process. We found an increase in the balanced accuracy score and F1-score to 0.835 and 0.856, respectively. The result indicates that the hybrid sampling method has the potential to be further developed to improve the classification model performance in various domain data sets.