Bridging Supervised and Generative Learning forExoplanet Discovery: The SS-GAN Model

Abstract

Exoplanets are a promising candidate for future space exploration projects. In this study, we present a semisupervised Generative Adversarial Network (SS-GAN), designed to detect and characterize exoplanetary transits from the Kepler space telescope. The proposed approach combines dropout regularization, batch normalization, label smoothing, and a composite validation strategy to enhance the stability and generalization capability of classical GANs. Using a comprehensive set of machine learning baselines, ranging from traditional classifiers such as Logistic Regression, Random Forest, and Support Vector Machine (SVM), to deep learning architectures including MLPs and residual networks, the proposed approach achieved an accuracy of 98.7% while effectively utilizing both labeled and unlabeled samples. It demonstrated robustness across multiple evaluation metrics, including balanced accuracy, F1-score, MCC, and threshold-free measures such as AUROC and AUPRC. These results underscore the importance of exploring generative and semisupervised architectures in astrophysics, paving the way for more efficient and reliable discovery of extraterrestrial candidates in upcoming large-scale surveys.