Physical Meaning of Features
This page explains what each family of AstroSpectro features measures physically, so that a SHAP ranking or a principal component can be read as astrophysics rather than as a list of column names. It is an interpretation page: the definitions of the features themselves are in the Feature Catalogue and the Feature Extraction Theory, and the importance ranking is in SHAP Interpretability. Statements about importance refer to the reference classifier run (G2, v6) and carry its leakage caveat.
The physical roles below are standard stellar-spectroscopy expectations, implemented by the line and band definitions in feature_engineering.py. Whether the model actually uses a feature for that reason is a separate question, addressed (partially) in Scientific Validation.
Families and their physical meaning
| Family | Typical features | What it traces | Why it discriminates A–M stars |
|---|---|---|---|
| Balmer lines | Hα, Hβ, Hγ, Hδ (equivalent width, FWHM, prominence, Hbeta_index) | Population of the n = 2 level of hydrogen, set mainly by temperature | Hydrogen lines are strongest in A-type stars and weaken towards both hotter and cooler types, so they separate A/F from G/K/M |
| Ca II H & K (3968 Å, 3934 Å) | CaIIK_eq_width, CaIIK_depth, CaIIK_prominence, K/Hβ ratio | Ionised calcium; strengthens towards cooler photospheres and also carries chromospheric and interstellar contributions | Strong, temperature-sensitive lines in F–K stars; their dependence on metallicity is expected but see the caution below |
| Mg b triplet (≈5170 Å) | Mg_b_eq_width, Mgb_index, individual Mg_5167/5173/5184 | Neutral magnesium lines, pressure-sensitive in G–K stars | Sensitive to temperature and, in cool stars, to surface gravity (dwarf/giant) |
| Na D doublet (≈5890 Å) | Na_D_eq_width, NaD_Lick | Neutral sodium; strong in cool photospheres, can be contaminated by interstellar absorption | Rises towards K and M types |
| Molecular bands | TiO (several bands), VO, CaH (CaH2, CaH3), CN, CH (G4300), MgH | Molecules that form only in cool atmospheres | TiO and VO dominate M-type spectra; CaH is a classical dwarf/giant indicator; the G band (CH) is strong in G–K stars |
| Continuum and colour | slope_blue/green/red/deep_red, cont_slope, cont_curvature, continuum_rms, colour indices, UV excess at 3900 Å | Overall spectral energy distribution, hence effective temperature (plus reddening and flux-calibration effects) | A smooth, robust temperature proxy; also the family most exposed to instrumental systematics |
The index feature_index_CaH2 (a CaH band near 6814–6846 Å, referenced to the continuum near 7042–7056 Å in the code) is labelled with the family "Ca II" in the SHAP discoveries.csv file. That label is a grouping artefact; physically it belongs with the molecular bands.
What the ranking suggests
In the SHAP analysis of the reference classifier run, the highest-ranked features are the Mg b equivalent width, the CaH2 index, the deep-red continuum slope, the Ca II K equivalent width and the Hα equivalent width, in that order; Na D, Hβ, G4300 and CN indices follow. Most of the top-30 are therefore line, band or continuum measurements that match the families above. The exact shares and ranks are on the SHAP Interpretability page.
This is consistent with the model using physically meaningful information, and it suggests that no single family explains the classification: line strengths, molecular bands and continuum slopes all appear among the leading features, as expected for a task that is largely a temperature sequence with gravity-sensitive refinements.
The same SHAP table also contains M_G, a Gaia-derived absolute magnitude, within its top 30, and colour features such as bp_g that are partly photometric; the ranking is therefore not a purely spectroscopic one.
Why some features rank higher than others
- Redundancy: several features measure related quantities (Mg b equivalent width, Mg b index, three individual Mg lines; many continuum slopes). Importance is shared or concentrated unpredictably among correlated features, so a feature's rank is not a clean measure of its physical weight.
- Dynamic range across classes: features that change monotonically from A to M (continuum slopes, molecular bands) give the model easy splits; features that peak in the middle of the sequence (Balmer lines) are informative but non-monotonic.
- Adjacent-class confusion: the largest remaining errors are between neighbouring types (F/G, G/K), where all features vary slowly; see F/G Confusion.
Caution: metallicity and Ca II
Ca II H&K is often cited as a metallicity indicator. In the scientific validation, the expected Ca II K–metallicity relation is not reproduced on the PHY-3500 · G1 · 43,019 spectra analysis (the correlation has the opposite sign to the expectation) and cannot be tested on the G2 set because metallicity is unavailable in spectro_only mode. The physical reading of Ca II features as metallicity tracers is therefore not supported by this work; see Scientific Validation.
Limitations of this interpretation
- Feature definitions use fixed wavelength windows; they do not correct for radial velocity or for blends.
- SHAP importance describes the trained model, not stellar physics: a feature can be important because it proxies another quantity.
- The Gaia-related columns in the reference run mean the ranking cannot be read as a purely spectroscopic one.