Using Machine Learning to determine the most important features in exoplanet verification

Abstract

Over a decade ago, NASA launched the Kepler Space Telescope in order to find earth-like planets revolving around sun-like stars in the hopes of finding habitable exoplanets. The Kepler pipeline picked up data for over 9000 astronomical bodies, out of which 52% were determined to be false positives, while the remaining 48% were candidates to be classified as exoplanets. The data collected from this mission can be used to assess and automatically classify Kepler Objects of Interest (KOIs) as exoplanets or false positives. Our goal in this work is to determine if some data features are more important than others in classifying an object as an exoplanet. To this end, we built 5 Machine Learning classification models (namely, logistic regression, support vector classifier, gradient boosting classifier, random forest classifier, and multilayer perceptron) and used 15 features to train and test them. We have included Machine Learning models that are explainable to help attain our goal, Our best predictor (random forests) achieved a prediction accuracy of 99% when evaluated with k-fold cross-validation. We evaluated the feature importances of our model and found that 5 of the features (Not Transit-like Flag, Centroid Offset Flag, Stellar Eclipse Flag, Ephemeris Match Indicate Contamination Flag, and Planetary Radius) out of the 15 selected ones make up roughly 75% of the overall feature importances. We hope that our findings can guide the selection of appropriate data to accurately predict exoplanet candidacy for future missions.