Uncertainty Estimation in Classification of Massive Stars Spectra using Conformal Predictions

Understanding and trusting the model's predictions have become fundamental, especially in critical and scientific contexts. Stellar spectra classification is essential in astrophysics, especially for massive stars. This work focuses on deep-learning models for classifying Hα spectral lines of massive stars into 21 classes and the conformal prediction approach for analyzing uncertainty. We built a deep neural network classifier using the ISOSCELES database. We used 342,780 spectral lines, and the experimental results showed an overall accuracy score of 0.92 and an average F1 score of 0.94. We comprehensively described the conformal prediction approach used to obtain the prediction sets. We generated calibration datasets of different sizes, computed the respective conformal scores, and used different error rates to evaluate diverse confidence levels. The small prediction sets showed confident predictions and demonstrated the model's effectiveness in maintaining high classification accuracy while providing reliable measures of uncertainty.