Classification of Poorly Time Sampled Light Curves of Periodic Variable Stars

Abstract

Classification of periodic variable light curves is important for scientific knowledge discovery and efficient use of telescopic resources for source follow-up. In practice, labeled light curves from catalogs with hundreds of flux measurements (the training set) may be used to classify curves from ongoing surveys with tens of flux measurements (the test set). Statistical classifiers generally assume that the probability of class given light curve features is the same for training and test sets. This assumption is unlikely to hold when the number of flux measurements per light curve varies widely between the two sets. We employ two methods to correct the problem—noisification and denoisification. With noisification we alter the training set to mimic the distribution of the test set and then construct a classifier on these altered data. With denoisification we construct a classifier on the well-sampled curves in the training set and probabilistically infer what poorly sampled curves in the test set would look like if we continued obtaining flux measurements. On periodic variable sources from a simulated data set and the OGLE survey, both of these methods outperform making no adjustments for training–test set differences.