Deep learning interpretability analysis for carbon star identification in <i>Gaia</i> DR3

Context. A large fraction of asymptotic giant branch (AGB) stars develop carbon-rich atmospheres during their evolution. Based on their color and luminosity, these carbon stars can easily be distinguished from many other kinds of stars. However, numerous G, K, and M giants also occupy the same region as carbon stars on the HR diagram. Despite this fact, their spectra exhibit differences, especially in the prominent CN molecular bands. Aims. We aim to distinguish carbon stars from other kinds of stars using Gaia ’s XP spectra while providing attributional interpretations of key features that are necessary for identification and even discovering new key spectral features. Methods. We propose a classification model named “GaiaNet”, an improved one-dimensional convolutional neural network specifically designed for handling Gaia ’s XP spectra. We utilized SHapley Additive exPlanations (SHAP), an approach for interpretability based on game theoretic, to determine SHAP values for each feature in a spectrum, enabling us to explain the output of the GaiaNet model and provide further meaningful analysis. Results. Compared to four traditional machine learning methods, the GaiaNet model exhibits an average classification accuracy improvement of approximately 0.3% on the validation set, with the highest accuracy reaching 100%. Utilizing the SHAP model, we present a clear spectroscopic heatmap highlighting molecular band absorption features primarily distributed around CN 773.3 and CN 895.0 , and we summarize five key feature regions for carbon star identification. Upon applying the trained classification model to the CSTAR sample with Gaia “xp_sampled_mean” spectra, we obtained 451 new candidate carbon stars as a by-product. Conclusions. Our algorithm is capable of discerning subtle feature differences from low-resolution spectra of Gaia , thereby assisting us in effectively identifying carbon stars with typically higher temperatures and weaker CN features while providing compelling attributive explanations. The interpretability analysis of deep learning holds significant potential in spectral identification.