Comparative Investigation of Machine Learning Model Performances for Exoplanet Transit Detection Using Kepler Data

Mankind has been fascinated with discovering new places, and exoplanet detection is no outlier. Exoplanets are planets that are located outside our solar system. The massive volume of data on these exoplanets, gathered by the National Aeronautics and Space Administration (NASA) missions like Kepler and Transiting Exoplanet Survey Satellite (TESS) on exoplanets, calls for automated classification. This investigation compares how each machine learning model performs for exoplanet detection using data gathered from the Kepler Space Telescope, focusing on three models: logistic regression, random forest, and shallow neural network. The performance metrics in this study are accuracy, precision, recall and F1 score.

The training, validation, and testing of all three models was performed on the same Kepler dataset. The models received four statistical features as input, derived from the star’s light curve (a graph of luminous flux against time): mean flux, standard deviation of the flux, maximum flux and minimum flux. The performance of each model was assessed using evaluation metrics of accuracy, precision, recall and F1 score derived from confusion matrices generated in Python.

The data on which the models were trained were derived from transit photometry, where periodic dips in the brightness of a star indicate an orbit. These dips can be seen in light curves.

This investigation demonstrates that machine learning can be effectively applied to exoplanet detection. This investigation also highlights the importance of model selection, feature representation and evaluation metrics in classification tasks.