Automatic Classification of Kepler Planetary Transit Candidates

In the first three years of operation the Kepler mission found 3,697 planet candidates from a set of 18,406 transit-like features detected on over 200,000 distinct stars. Vetting candidate signals manually by inspecting light curves and other diagnostic information is a labor intensive effort. Additionally, this classification methodology does not yield any information about the quality of planet candidates; all candidates are as credible as any other candidate. The torrent of exoplanet discoveries will continue after Kepler as there will be a number of exoplanet surveys that have an even broader search area. This paper presents the application of machine-learning techniques to the classification of exoplanet transit-like signals present in the \Kepler light curve data. Transit-like detections are transformed into a uniform set of real-numbered attributes, the most important of which are described in this paper. Each of the known transit-like detections is assigned a class of planet candidate; astrophysical false positive; or systematic, instrumental noise. We use a random forest algorithm to learn the mapping from attributes to classes on this training set. The random forest algorithm has been used previously to classify variable stars; this is the first time it has been used for exoplanet classification. We are able to achieve an overall error rate of 5.85% and an error rate for classifying exoplanets candidates of 2.81%.