Machine Learning Approaches for Exoplanet Detection and Classification: A Comparative Study of Traditional and Deep Learning Models with Explainable AI

The exponential surge in stellar photometric data from missions like Kepler and TESS has created a bottleneck in exoplanet discovery; while manual vetting is unscalable, automated systems must prioritize scientific reliability over raw accuracy. Deep Learning (DL) models are often proposed as universal solutions, yet their “black box” nature and tendency to fail on rare events pose significant risks for astrophysical discovery. This research presents a novel comparative analysis of domain-informed classical machine learning versus end-to-end deep learning, specifically targeting the challenge of extreme class imbalance (136:1). We evaluate an XGBoost classifier with 12 engineered statistical features against a 1D-CNN trained on raw light curves, utilizing Grad-CAM and SHAP for interpretability. Using a curated dataset of 5,087 Kepler Objects of Interest, results revealed a critical “accuracy paradox”. While CNNs achieved a higher overall accuracy of 96.15%, they catastrophically failed to detect the minority class (1.71% precision for confirmed planets). Conversely, XGBoost demonstrated superior scientific utility with 87% precision and 0.87 F1-score. Explainability analysis exposed that CNNs frequently overfit to background noise, whereas XGBoost relied on physically meaningful features like transit skewness. These findings underscore that domain knowledge remains superior to algorithmic complexity for rare-event detection, offering a robust framework for future automated vetting in large-scale astronomical surveys.