Physical Purity & Leakage Prevention
AstroSpectro distinguishes between physical features (spectroscopic properties of stars) and non-physical features (observational, instrumental, or Gaia-derived metadata).
Types of problematic features
1. Positional features: ra, dec
LAMOST observes targets according to survey programs. Certain programs target specific spectral types in specific sky regions. Including ra and dec allows the model to exploit this geographic correlation rather than learning spectral physics.
Effect: Removing ra/dec increased BA from 84% → 87%.
2. Instrumental metadata: fiber_id, seeing, jd
These encode nothing about the stellar physics — only the observational conditions.
3. Observational proxies: redshift, snr_*
For Galactic stars, redshift ≈ radial velocity / c ≈ 0. It is correlated with spectral class only because of population dynamics, not intrinsic spectral physics. SNR bands can correlate with spectral type due to instrumental sensitivity curves.
4. Gaia-derived features: teff_gspphot, logg_gspphot
Gaia's GSP-Phot derives teff_gspphot from BP/RP photometry and spectra. Using this to predict spectral class (which is itself a temperature sequence) creates a partially circular inference chain. These features are excluded in spectro_only=True mode.
Rule of thumb
A feature is physical if a spectroscopist working with only the spectrum could compute it.