Evaluating machine learning models for stellar spectral classification using Gaia DR3

This work evaluates the machine learning models for stellar spectral classification based on the third data release (DR3) of Gaia. We have examined how different machine learning models and feature selection techniques impact the classification accuracy. We have used seven supervised machine learning algorithms (Decision Tree, k-Nearest Neighbour, Naive Bayes classifier, Artificial Neural Networks, Random Forest and Support Vector Machine) for performing Morgan-Keenan spectral classification of A, F, G, K and M type stars. For feature selection, we used four different methods (Mutual Information, chi 2, F-test and Pearson Correlation). The Mutual Information feature selection method gave the best performance with an average accuracy of 88.76% across all models. The Artificial Neural Networks classifier showed the highest average accuracy of 90.97% across the four feature selection methods. The combination of Mutual Information feature selection and Artificial Neural Network has given the best classification accuracy of 91.43%. The four feature selection methods identified ten common features (RAICRS (deg), G (mag), BP (mag), log g (cgs), Teff (K), R (R circle dot), M (M circle dot), t (Gyr), z (km s-1) and Evol) that dominate the spectral classification. We discuss the implication of these selected features based on our understanding of astrophysical parameters associated with various spectral classes. Based on our review of the literature, this appears to be the first detailed and robust empirical study with Gaia DR3.