Detecting Label Noise in Kepler Confirmed Planet Catalog using Machine Learning

Catalogs of validated planets are built based on a myriad of validation systems, some more automated than others. These systems differ in the types of data and catalogs they use and in their model construction process. The automated systems make use of statistical or machine learning tools to validate new planets that are added to the pool of previously confirmed planets, but no insight or analysis is performed on corroborating or demoting those already in the pool. There might be no consensus among these tools for the validated exoplanets; one model could be very confident that a given transit signal is a planet, and another could be very confident it is not. Although the number of such cases is very small, label noise must be studied and detected to improve the accuracy of the lists of confirmed planets and false positives that are later used for studies about planet occurrence rates and for creating catalogs built upon this population. In this talk, we discuss our efforts in studying and detecting mislabeled cases in the Kepler Cumulative KOI catalog using machine learning tools specially designed for label noise detection.