Identification of Discriminative Features from Light Curves for Automatic Classification of Variable Stars

Variable stars are stars whose brightness changes overtime. Due to their change of brightness, they are relatively easy to observe. Astronomers use variable stars as a tool to learn about the formation and evolution of the system that they are in. Different types of variable stars provide unique information about the host system. To classify the types of these stars, astronomers traditionally look at their light curve to see how their light changes over time. Recently, observational data of variable stars has increased exponentially, making machine learning based classification a viable alternative to manual classification. In this work, we tackle the task of extracting and selecting a good set of features from light curve profiles retrieved from ASAS-SN archive for variable star classification. We found that by combining several feature selection methods in an increasing order of their aggressiveness towards feature reduction, we obtained a set of highly discriminative features which is smaller in size as compared to that of Mutual Information, Gradient Boosted Tree, L1 Regularization, Elastic net, and manually chosen by the expert, while still maintaining comparable classification performance.