Title:TSARDI: a Machine Learning data rejection algorithm for transiting exoplanet light curves

Abstract:We present TSARDI, an efficient rejection algorithm designed to improve the transit detection efficiency in data collected by large scale surveys. TSARDI is based on the Machine Learning clustering algorithm DBSCAN, and its purpose is to serve as a robust and adaptable filter aiming to identify unwanted noise points left over from data detrending processes. TSARDI is an unsupervised method, which can treat each light curve individually; there is no need of previous knowledge of any other field light curves. We conduct a simulated transit search by injecting planets on real data obtained by the QES project and show that TSARDI leads to an overall transit detection efficiency increase of $\sim$11\%, compared to results obtained from the same sample, but using a standard sigma-clip algorithm. For the brighter end of our sample (host star magnitude < 12), TSARDI achieves a detection efficiency of $\sim$80\% of injected planets. While our algorithm has been developed primarily for the field of exoplanets, it is easily adaptable and extendable for use in any time series.