Improving Imbalanced Dataset Classification Using Oversampling and Gradient Boosting

摘要

Imbalanced data classification is challenging task for various datasets in the real world. One of technique to enlarge the sample in minority class is oversampling to fix size as majority class. This research aims to test SMOTE, Borderline-SMOTE, and ADASYN to handle dataset imbalance and to observe its impact toward classification accuracy. Gradient Boosting applied as a classifier and seven datasets are used in this research. Accuracy, recall, precision, F1-Score, AUC were also implemented to measure classifier performance. Experiments showed that oversampling technic increase accuracy from 2% to 11% for the dataset Mammography, Liver Disorders, Diabetes (Pima Indian), Indian Liver, Habberman, and Immunotherapy. Borderline-SMOTE increases higher accuracy compared to other oversampling method. Surprisingly, Breast Cancer Wisconsin has steady accuracy with or without oversampling. Even though, oversampling good for data imbalanced, the sensibility of oversampling algorithm and the nature of dataset must considered.