How effective oversampling techniques are in classifying potentially hazardous asteroids

Abstract

Potentially hazardous asteroids require strict surveillance to ensure the safety of our planet. However, the vast amount of increasing astronomical data makes it troublesome for humans to study these asteroids. Hence, machine learning techniques are used to classify these hazardous asteroids. However, machine learning models are not robust for distinguishing imbalanced classes. Various undersampling and oversampling techniques are used to address this problem. In our study, we refrained from using any undersampling technique as we did not want to lose any valuable information. Instead, we employed various oversampling techniques, including random Oversampling, SMOTE (Synthetic minority over-sampling technique), ADASYN (Adaptive Synthetic Sampling), BorderlineSmote, KMeansSmote, and SVMSmote. For each oversampling technique, we trained the Random Forest, XGBoost, LightGBM, HistGradientBoosting, and AdaBoost classifiers. Our research presents a detailed study of these oversampling techniques to determine which one is more suitable for our dataset.