Protostellar classification using supervised machine learning algorithms

TL;DR

This study applies various supervised machine learning algorithms to classify young stellar objects in Orion based on multiwavelength flux data, achieving up to 82% accuracy with gradient boosting machines, demonstrating machine learning's potential in astronomical classification tasks.

Contribution

The paper evaluates and compares multiple machine learning algorithms for YSO classification, highlighting the effectiveness of gradient boosting machines and the importance of specific flux features.

Findings

Gradient boosting machine achieved 82% accuracy.

Including 3.6 μm and 24 μm fluxes improves classification.

Machine learning offers a rapid classification method for large astronomical datasets.

Abstract

Classification of young stellar objects (YSOs) into different evolutionary stages helps us to understand the formation process of new stars and planetary systems. Such classification has traditionally been based on spectral energy distributions (SEDs). An alternative approach is provided by supervised machine learning algorithms. We attempt to classify a sample of Orion YSOs into different classes, where each source has already been classified using multiwavelength SED analysis. We used 8 different learning algorithms to classify the target YSOs, namely a decision tree, random forest, gradient boosting machine (GBM), logistic regression, na\"ive Bayes classifier, $k$-nearest neighbour classifier, support vector machine, and neural network. The classifiers were trained and tested by using a 10-fold cross-validation procedure. As the learning features, we employed ten continuum flux densities spanning from the near-IR to submm. With a classification accuracy of 82% (with respect to the SED-based classes), a GBM algorithm was found to exhibit the best performance. The lowest accuracy of 47% was obtained with a na\"ive Bayes classifier. Our analysis suggests that the inclusion of the 3.6 $\mu$m and 24 $\mu$m flux densities is useful to maximise the YSO classification accuracy. Although machine learning has the potential to provide a rapid and fairly reliable way to classify YSOs, an SED analysis is still needed to derive the physical properties of the sources, and to create the labelled training data. The classification accuracies can be improved with respect to the present results by using larger data sets, more detailed missing value imputation, and advanced ensemble methods. Overall, the application of machine learning is expected to be very useful in the era of big astronomical data, for example to quickly assemble interesting target source samples for follow-up studies.