CLUSTERING-BASED FEATURE LEARNING ON VARIABLE STARS

Abstract

The success of automatic lightcurve classification using machine learning techniques depends strongly of the lightcurve representation. Usually, lightcurves are represented as a vector of many statistical descriptors chosen arbitrarily by astronomers. These descriptors commonly demand a lot of computing power to calculate and do not in any way guarantee a good performance on the final classification task. In recent years, a new way of obtaining descriptors for data like natural images and sound has been developed, called feature learning. Feature learning uses data to learn a model which then can be used to encode the data to a new representation suitable for machine learning tasks such as automatic classification. We present an application of feature learning on lightcurve data based on Affinity Propagation clustering and the Time Warp Edit Distance similarity measure for time series, which has allowed us to represent lightcurves without the computational burden of fitting statistical models and finding descriptors for the data.