Hierarchical Classification of Variable Stars Using Neural Networks

Variable stars play a prominent role in our study of the universe and are essential to estimating cosmological parameters. They are considered '‘standard candles’' due to their intrinsic variability, which allows their distances to be calculated. With the proliferation of large-scale sky surveys that generate over 20 Terabytes of light-curve observations every day, automated methods are necessary to reduce manual efforts when classifying variable stars. To automate such classification, astronomers have developed various machine learning algorithms. Existing algorithms exploit star properties but fail to use the hierarchical structure known to exist in a specific family of stars. We believe embedding hierarchical information of stars into a learning algorithm can lead to more robust and efficient machine learning models. The goal of this thesis is to explore various approaches that exploit the hierarchical structure of stars within a neural network architecture. Results show the conditions under which adding information of the intrinsic hierarchical structure helps increase generalization performance.Computer Science, Department o