Theoretical foundations of spectroscopic feature extraction in AstroSpectro: the Physics-Informed approach, the 6 descriptor families, spectro_only mode, and a SHAP check of how much of the model's attribution comes from physical features.
Complete guide to training a spectral classifier with SpectralClassifier: available models (XGBoost, LightGBM, CatBoost, ExtraTrees…), GridSearchCV and Optuna tuning, W&B tracking, spectro_only mode, session artifacts.
AstroSpectro's most important metrics and scientific findings.
AstroSpectro results in two separate studies: unsupervised dimensionality reduction (PCA, UMAP, t-SNE, HDBSCAN) on the PHY-3500 dataset (G1, 43,019 spectra), and the reference XGBoost classifier with its SHAP attribution on the 350k-spectrum generation (G2, run v6, which includes Gaia/photometric features).
SHAP analysis of the AstroSpectro XGBoost classifier — the central scientific finding.
Results of the reference XGBoost classifier (run v6, 350k-spectrum generation G2, five types A–M): balanced accuracy, macro F1 and ROC-AUC from runs.json, per-class performance, SHAP attribution — with the Gaia/photometric features it contains — and its qualitative link with the PHY-3500 UMAP study (G1).