Positive Unlabelled Learning to discover hybrid pulsating stars in the TESS catalogue

Existence of stellar pulsations allows to use asteroseismology to learn about the interiors of stars that otherwise could not be discovered with classic observational methods. Hybrid pulsating stars that exhibit both pressure and gravity modes are of utmost astrophysical interest owing to the fact that joint interpretation of p and g modes in these stars tells us about their interior physics across the entire mass coordinate. Hybrid pulsators amongst OBAF-type stars are relatively rare. Automatic detection of rare stars in large astronomical surveys faces significant challenges of manual labelling for supervised and low interpretability for unsupervised machine learning methods. In this study, we propose Positive Unlabelled (PU) Learning as a semi-supervised learning setting for the identification of hybrid pulsators, which are crucial from the perspective of stellar structure and evolution studies. We work with QLP light curves, from which we remove instrumental noise and extract astrophysical frequencies. Next, we use PU-bagging to train an ensemble of decision trees (DTs). Each DT is trained using the entire positive and a bootstrapped sample of unlabelled light curves from a single TESS sector. New positive candidates are found by selecting the unlabelled instances with the highest average out-of-bag scores. We validate the model by injecting the unlabelled set with known labelled non-hybrids and establish a probabilistic threshold on the final probability distribution.