Current Status
This page summarises where AstroSpectro stands: which runs are the current references, what is known to be wrong or incomplete, and what is planned. It is updated whenever the reference run changes.
Reference runs
| Analysis | Current reference | Data generation |
|---|---|---|
| Spectral-type classifier | ST_350k_xgb_AFGKM_v6 (saved 2026-05-22T16:12:36Z) | G2 — 349,998 spectra processed |
| SHAP attribution | SHAP of the reference classifier | G2 |
| Scientific validation | Feature-versus-theory checks | G2 |
| PCA, UMAP, t-SNE, HDBSCAN, autoencoder | PHY-3500 runs | PHY-3500 · G1 · 43,019 spectra |
Metrics: Classification Performance · Model Card · Key Results.
Known issues
Gaia/photometric features in the reference classifier
Run v6 (ST_350k_xgb_AFGKM_v6) was trained with spectro_only enabled, but its 99 selected features include 4 Gaia/photometric columns (M_G, M_BP, M_RP, ag_gspphot). Its metrics are therefore not purely spectroscopic; see the Data Leakage Audit.
The cause was in the code: the spectro_only exclusion worked by column-name prefixes, and the Gaia-derived columns did not match any of them (see Classification). Fixed in the code with an explicit list of Gaia/photometric columns (the 22 columns absent from v7). v7 was trained with this fix, but no complete spectro-only reference run exists yet.
No complete spectro-only reference run yet
The fix is in the code, and the validation run v7 (VALIDATION_spectro_only_v7) was trained with it, without the Gaia/photometric columns. That run is partial: according to the author, its Gaia cross-match is incomplete and its SHAP analysis was not completed. It is therefore not used as a reference and its metrics are not reported as results. No complete spectro-only reference run exists yet, so the reference classifier is still v6, with its Gaia columns.
Two data generations
The dimensionality-reduction pages and the classifier use different datasets (G1 and G2). The dimensionality-reduction runs on G2 exist but are not published yet; until they are, results from the two studies must not be combined. See Data & Datasets.
Other open points
| Area | Issue | Page |
|---|---|---|
| Validation | Hα FWHM test fails; Ca II K versus metallicity is inverted on G1 and untestable on G2 | Scientific Validation |
| Preprocessing | Inverse variance not rescaled after median normalisation | Preprocessing |
| Labels | F/G errors cannot be separated from label noise without an external reference | F/G Confusion |
| Data provenance | Gaia match rate and per-filter counts are not in the exported fact pack | Data & Datasets |
Planned work
These items are intentions; none of them has published results yet.
- A complete spectroscopic-only reference run — retrain with the corrected exclusion list on the full generation, with a complete Gaia cross-match and a full SHAP analysis, and use it to replace v6 as the reference.
- Hyperparameter optimisation — a systematic search (Optuna is the planned tool) instead of reusing the parameters of an earlier run.
- Dimensionality reduction on G2 — decide whether the PHY-3500 study is extended to the larger generation.
- Publication — a short manuscript on the classifier is in preparation; a complete spectro-only reference run (the code fix is in place) is a prerequisite.
See the Roadmap and Limitations.