Penggunaan Smote Untuk Meningkatkan Kinerja Model Deteksi Eksoplanet Berdasarkan Data Fluks Bintang

Abstract

detecting them is a major challenge in modern astronomy. One of the primary methods for detecting exoplanets is the transit method, which takes advantage of the changes in a star's light flux as a planet passes in front of it. However, the dataset obtained from NASA's Campaign 3 mission shows significant class imbalance, where the number of stars without exoplanets far exceeds those with exoplanets. In this study, a Convolutional Neural Network (CNN) method is used to detect exoplanets from the imbalanced stellar flux dataset. To address this imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) is applied, aiming to enhance the representation of the minority class—exoplanets—without altering the proportion of the majority class. This research aims to explore how SMOTE can resolve data imbalance issues and assess its impact on CNN model accuracy in detecting exoplanets. The model is developed and tested with and without the application of SMOTE, and the performance results are compared. In the model without SMOTE, the training accuracy reached 93.5%, and the testing accuracy was 99.12%, though the model exhibited bias toward the majority class. With SMOTE applied, the training accuracy reached 99.2%, and the testing accuracy was 98.6%. Although the overall accuracy slightly decreased, the SMOTE-applied model showed improvement in detecting exoplanets, as evidenced by better precision, recall, and F1-score metrics. The application of SMOTE successfully addressed data imbalance and improved the CNN model's performance in detecting exoplanets, particularly in identifying the minority class. This study contributes to the development of machine learning-based exoplanet detection methods, which can serve as a reference for future research in the exploration of planets beyond the solar system.