Improving Astronomical Time Series Classifiation via Data Augmentation

Advancements in survey telescopes have revolutionized our ability to observe time-varying objects in the sky. These objects are represented by graphs known as light curves that show their brightness over a period of time. These light curves need to be classified rapidly as millions of observations can be made in a single night. However, the performance of existing supervised machine learning classifiers is limited as real-world astronomical datasets often have highly imbalanced class distributions. To tackle this issue, data augmentation techniques can be applied to transform small imbalanced datasets into large balanced datasets, thereby improving the accuracy of these classification algorithms.

 

My research is primarily focused on a data augmentation methodology for astronomical time series data using Generative Adversarial Networks. The approach involves designing a GAN capable of performing conditional generation based on the class of the light curve and other physical parameters such as observation time and amplitude. The data generated is then used to train a classifier to yield better results.

 

The implementation uses an auxiliary classifier GAN architecture with the Wasserstein gradient penalty loss function, where the generative and discriminative steps are implemented with deconvolutional and convolutional networks. Generated light curves from the GAN are fed to a classifier that consists of an ensemble of convolutional neural networks (CNN) to yield better classification results. 

 

As on date of writing this, I am still in the initial stages of this project and will continue to work in order to bring this research project to fruition.