Optimized sampling of SDSS-IV MaStar spectra for stellar classification using supervised models

TL;DR

This paper demonstrates that active learning algorithms can efficiently select the most informative stellar spectra from the MaStar library, reducing labeling effort while improving classification performance for stellar parameters.

Contribution

It applies active learning to astronomical spectral data, showing its effectiveness in class-imbalanced datasets for stellar classification tasks.

Findings

Active learning outperforms random sampling in classification metrics.

Fewer training samples are needed to achieve high performance.

Active learning can optimize target selection for follow-up observations.

Abstract

Supervised machine learning models are increasingly being used for solving the problem of stellar classification of spectroscopic data. However, training such models requires a large number of labelled instances, the collection of which is usually costly in both time and expertise. Active learning algorithms minimize training dataset sizes by keeping only the most informative instances. This paper explores the application of active learning to sampling stellar spectra using data from a highly class-imbalanced dataset. We utilize the MaStar library from the SDSS DR17 along with its associated stellar parameter catalogue. A preprocessing pipeline that includes feature selection, scaling, and dimensionality reduction is applied to the data. Using different active learning algorithms, we iteratively query instances, where the model or committee of models exhibits the highest uncertainty or disagreement, respectively. We assess the effectiveness of the sampling techniques by comparing several performance metrics of supervised-learning models trained on the queried samples with randomly-sampled counterparts. Evaluation metrics include specificity, sensitivity, and the area under the curve; in addition to the Matthew's correlation coefficient, which accounts for class imbalance. We apply this procedure to effective temperature, surface gravity, and iron metallicity, separately. Our results demonstrate the effectiveness of active learning algorithms in selecting samples that produce performance metrics superior to random sampling and even stratified samples, with fewer training instances. Active learning is recommended for prioritizing instance labelling of astronomical-survey data by experts or crowdsourcing to mitigate the high time cost. Its effectiveness can be further exploited in selection of targets for follow-up observations in automated astronomical surveys.