Machine Learning-based Classification of Variable Stars

Abstract The classification of variable stars, essential for revealing information on stellar properties and cosmic distances, traditionally relied on statistical methods and limited data. With the emergence of transformative machine learning methodologies and large stellar surveys, we are able to perform more efficient, accurate, and robust handling of expansive databases. We deploy two different machine learning models—random forest and XGBoost—that effectively classify variable stars identified in the most recent phase of the Optical Gravitational Lensing Experiment using four features: (1) time of minimum brightness, (2) I -band amplitude, (3) mean or maximum I -band magnitude, and (4) period. The two models achieve a cross-correlation score of 99.00%, indicating largely shared classifications. The results produced by the models uncover new insights into the precision of supervised learning models in predicting variable stars with the aid of the most up-to-date survey data.