Bridging the Black Box: Data Mining Enhanced Machine Learning Interpretability for Exoplanet Discovery

The widely accepted notion that there is an inherent trade-off between accuracy and interpretability in artificial intelligence (AI) models is often misleading. While black-box machine learning (ML) models are commonly viewed as essential for achieving high predictive and classification performance, this paper demonstrates that deterministic data mining (DM) techniques can enhance ML interpretability without compromising accuracy or efficiency. Using the Periodic Motion Detection (PMD) algorithm, we show that deterministic methods can improve transparency in ML decision-making processes without affecting computational time or sacrificing accuracy and precision, particularly in large, noisy real-world datasets. To validate its performance, the PMD process was applied to 4,000 light curves from the Kepler Space Telescope, each containing 15,000 data points, to extract known exoplanet signals. Its performance was compared to that of eleven ML models. After normalizing results for GPU computation, PMD processed a single light curve in just 0.021 seconds—at least an order of magnitude faster than all ML models tested. Additionally, it achieved an accuracy of 93.23%, precision of 98.76%, recall of 87.55%, and specificity of 98.9%, matching or exceeding the performance of the ML models. PMD requires minimal preprocessing and no iterative training, maintaining low design complexity and computational costs. Rather than replacing ML, this study emphasizes how deterministic methods can serve as an efficient interpretability layer for black-box models. By integrating DM techniques like PMD, AI practitioners can gain valuable insights into black box decision processes while preserving the predictive power of machine learning.