Leaves on trees: identifying halo stars with extreme gradient boosted trees

TL;DR

This paper develops a machine learning classifier using Gradient Boosted Trees to identify halo stars in Gaia data, achieving high accuracy even with limited phase-space information, aiding understanding of galaxy formation.

Contribution

The study introduces a novel supervised classifier trained on Gaia simulations and catalog cross-matches to reliably identify halo stars with limited data.

Findings

Achieves 90% recovery of halo stars with full phase-space data

Detects 337 high-confidence halo stars in TGAS data

Performance degrades with large parallax errors

Abstract

Extended stellar haloes are a natural by-product of the hierarchical formation of massive galaxies. If merging is a non-negligible factor in the growth of our Galaxy, evidence of such events should be encoded in its stellar halo. Reliable identification of genuine halo stars is a challenging task however. The 1st Gaia data release contains the positions, parallaxes and proper motions for over 2 million stars, mostly in the Solar neighbourhood. Gaia DR2 will enlarge this sample to over 1.5 billion stars, the brightest ~5 million of which will have a full phase-space information. Our aim is to develop a machine learning model to reliably identify halo stars, even when their full phase-space information is not available. We use the Gradient Boosted Trees algorithm to build a supervised halo star classifier. The classifier is trained on a sample extracted from the Gaia Universe Model Snapshot, convolved with the errors of TGAS, as well as with the expected uncertainties of the upcoming Gaia DR2. We also trained our classifier on the cross-match between the TGAS and RAVE catalogues, where the halo stars are labelled in an entirely model independent way. We then use this model to identify halo stars in TGAS. When full phase- space information is available and for Gaia DR2-like uncertainties, our classifier is able to recover 90% of the halo stars with at most 30% distance errors, in a completely unseen test set, and with negligible levels of contamination. When line-of-sight velocity is not available, we recover ~60% of such halo stars, with less than 10% contamination. When applied to the TGAS data, our classifier detects 337 high confidence RGB halo stars. Although small, this number is consistent with the expectation from models given the data uncertainties. The large parallax errors are the biggest limitation to identify a larger number of halo stars in all the cases studied.