Effective detection of variable celestial objects using machine learning-based periodic analysis

This paper tackles the problem of effectively detecting variable celestial objects whose brightness periodically changes over time. This problem is crucial in studying the evolution and structure of the universe and elucidating physical phenomena. The method by Sesar et al. is one of the popular approaches used in detecting variable celestial objects that uses statistical data of celestial time series, such as intrinsic variability $\sigma$ and ${\chi }^2$, etc. However, since statistical data is an aggregation of celestial time series, the previous approaches do not take advantage of the periodicity, which is the inherent characteristic of variable celestial objects; it fails to find variable celestial objects effectively. To solve such a problem, we propose an approach to detecting variable celestial objects using periodic analysis. Our approach uses sparse modeling as periodic analysis since celestial time series is typically sparse and sparse modeling can effectively obtain periodicities of the celestial objects from sparse time series. By exploiting the periodicities of the celestial objects as features, we perform binary classification to estimate whether a celestial object is a variable celestial object. To show the effectiveness of our approach, we evaluated our approach using Hyper SuprimeCam (HSC) PDR2 dataset, and we confirmed that AUC of our approach is 0.939 while AUC of the previous approach is 0.750; our approach can more effectively detect variable celestial objects.