A New Approach To Domain Adaptation Applied To Supernova Photometric Classification

Supernova Type Ia plays a vital role in the measurement of the cosmological parameters. It is used as ‘standard candles’ for measuring extragalactic distances. There are other types of supernovae like Supernova Type Ib and Ic that closely resemble Supernova Type Ia (but are not as useful as Supernova Type Ia). Large telescopic surveys capture light curves of these supernovae events referred as photometric observations, which include all the three types. Thus, accurate classification of supernovae from these photometric observations is desirable for proper calculation of cosmological parameters. The existing method for classification of supernova photometric observations is based on spectroscopic method, which is very cumbersome and expensive. In the future, with the increase in photometric surveys, myriad number of supernova photometric observations is expected. Thus, an efficient method for the classification of supernovae is required to replace existing methods. We also want to take advantage of existing dataset classified by spectroscopic method for the classification of upcoming photometric dataset. Since, these two datasets belong to different domains, an adaptive mechanism across the domains is required. Thus, we propose a method to generate a predictive model using domain adaptation with active learning that will classify supernovae (Ia, Ib, Ic) using spectroscopic data (aka source data) as a training set and photometric data (aka target data) as a testing set. Our method includes two concepts of machine learning: 1. Domain adaptation technique is used to transfer the source domain information to the target domain. 2. Active-learning technique is used to rely on only few target domain labels in a non-uniform distribution to build an effective model. The experiments and results show that our method outperforms various domain adaptation techniques with significant increase in classification accuracy.