VVV Ensemble Learning for Variable-star Extraction and Tagging (VELVET): Robust Classification of 5.2Million Variable Stars in the VVV Survey Using XGBoost

In this work, we applied the machine learning technique XGBoost to classify ~5.2 million variable stars in the VVV survey. Using a training set cross-matched from OGLE-IV, Gaia DR3, and SPICY catalogs, we build an XGBoost classifier on a curated set of photometric and time-series parameters. We validate the performance of XGBoost using a 20-fold internal cross-validation to demonstrate the robustness of the classifier. Using the resulting classification metrics, we recover a high confidence reference sample of ~24,000 RR Lyrae stars in the VVV survey, demonstrating the effectiveness of the classification method and the use of the resulting catalog for stellar population studies. We extend this analysis to other variable star types to show how the methodology is broadly applicable. Both the full classification catalog and the source code are publicly available. This methodology can be extended to other variable star types and surveys to produce robust classifications for large datasets of variable stars and other time-domain sources.