Classification of Astronomical Spectra Using PCA-Compressed Flux and Inverse-Variance Features

This paper evaluates a signal-processing and supervised-learning pipeline for classifying SDSS DR17 spectra into stars, galaxies, and quasars. Flux and inverse-variance vectors are standardized, separately compressed by PCA, and used by 12 classifiers. Small MLP and Hist. GBDT tied for the highest F+I accuracy (93.9%), while Small MLP achieved the highest macro-F1 (92.1%) and LightGBM the highest balanced accuracy (91.9%). Ten repeated stratified train/validation splits with a fixed test set showed low variability among the leading models, while LightGBM benefited most clearly from inverse-variance features.