| Summary: | Transit photometry is a widely used method for searching exoplanets. For example, NASA’s Transiting Exoplanet Survey Satellite (TESS) utilizes this technique. However, identifying exoplanet candidates requires significant human effort to process light curves; a workflow with minimal human input is desirable. Unfortunately, very few labeled training data are available (i.e., light curves labeled as planet candidates), which makes automatic classification di cult. Here, we propose a new approach to identify planet candidates using relevance-feedback accelerated few-shot learning. We generate many labeled synthetic light curves with and without transits by combining a simple physics-based transit injection model with a statistics-based generative model seeded with abundant non-transiting (“noise”) light curve data. After comparing multiple methods, we selected and trained a generic XGBoost classifier offline on only unfolded and diffused synthetic light curves. We adapted it online by feeding back a few observed and misclassified light curves. The result is an exoplanet classifier with the currently best-known recall and precision.
|
|---|