Likely Planet Candidates Identified by Machine Learning Applied to Four Years of Kepler Data
Abstract
Over 3,200 transiting planet candidates, 134 confirmed planets, and ~2,400 eclipsing binaries have been identified by the Kepler Science pipeline since launch in March 2009. Compiling the list of candidates is an intensive manual effort as over 18,000 transit-like signatures are identified for a run across 34 months. The vast majority are caused by artifacts that mimic transits. While the pipeline provides diagnostics that can reduce the initial list down to ~5,000 light curves, this effort can overlook valid planetary candidates. The large number of diagnostics 100) makes it difficult to examine all the information available in identifying planetary candidates. The effort required for vetting all threshold-crossing events (TCEs) takes several months by many individuals associated with the Kepler Threshold Crossing Event Review Team (TCERT). We have developed a random-forest classifier that decides whether a TCE should be called `planet candidate’, `astrophysical false positive’, or `non-transiting phenomena’. Ideally a machine learning algorithm will generate a list of candidates that approximates those generated by human review, thereby allowing the humans to focus on the most interesting cases. By using a machine learning-based auto-vetting process, we have the opportunity to identify the most important metrics and diagnostics for separating signatures of transiting planets and eclipsing binaries from instrument-induced features, thereby improving the efficiency of the manual effort. We report the results of a applying a random forest classifier to four years of Kepler data. We present characteristics of the likely planet candidates identified by the auto-vetter as well as those objects classified as astrophysical false positives (eclipsing binaries and background eclipsing binaries). We examine the auto-vetter's performance through receiver operating characteristic curves for each of three classes: planet candidate, astrophysical false positive, and artifact. Funding for this mission is provided by NASA’s Science Mission Directorate.
- Publication:
-
AAS/Division for Planetary Sciences Meeting Abstracts #45
- Pub Date:
- October 2013
- Bibcode:
- 2013DPS....4520406J