A machine-learning photometric classifier for massive stars in nearby galaxies I. The method

TL;DR

This paper presents a machine learning photometric classification method for massive stars in nearby galaxies, achieving high accuracy in identifying different stellar types despite challenges like class imbalance and metallicity effects.

Contribution

It introduces an ensemble machine learning approach combining multiple classifiers and synthetic data generation to classify massive stars across different galaxies with improved accuracy.

Findings

Overall weighted accuracy of ~83% in classification.

High recovery rate (~94%) for Red supergiants.

Lower accuracy (~30-45%) for rare classes like LBVs and Wolf-Rayets.

Abstract

(abridged) Mass loss is a key parameter in the evolution of massive stars, with discrepancies between theory and observations and with unknown importance of the episodic mass loss. To address this we need increased numbers of classified sources stars spanning a range of metallicity environments. We aim to remedy the situation by applying machine learning techniques to recently available extensive photometric catalogs. We used IR/Spitzer and optical/Pan-STARRS, with Gaia astrometric information, to compile a large catalog of known massive stars in M31 and M33, which were grouped in Blue, Red, Yellow, B[e] supergiants, Luminous Blue Variables, Wolf-Rayet, and background galaxies. Due to the high imbalance, we implemented synthetic data generation to populate the underrepresented classes and improve separation by undersampling the majority class. We built an ensemble classifier using color indices. The probabilities from Support Vector Classification, Random Forests, and Multi-layer Perceptron were combined for the final classification. The overall weighted balanced accuracy is ~83%, recovering Red supergiants at ~94%, Blue/Yellow/B[e] supergiants and background galaxies at ~50-80%, Wolf-Rayets at ~45%, and Luminous Blue Variables at ~30%, mainly due to their small sample sizes. The mixing of spectral types (no strict boundaries in their color indices) complicates the classification. Independent application to IC 1613, WLM, and Sextans A galaxies resulted in an overall lower accuracy of ~70%, attributed to metallicity and extinction effects. The missing data imputation was explored using simple replacement with mean values and an iterative imputor, which proved more capable. We also found that r-i and y-[3.6] were the most important features. Our method, although limited by the sampling of the feature space, is efficient in classifying sources with missing data and at lower metallicitites.