Hierarchical stellar refinement preserves measured recall and reduces contaminant selections
5 Oct 2026 · embeddings
Can finer unlabelled source families retain the useful coarse stellar anchor while improving discrimination?
Two frozen coarse parents with weighted two-child masked Gaussian refinement; 3334 training identities only, no class loss. Preserve coarse probabilities and identity fallbacks; original calibration-only interpretation and thresholds, both variants reported.
Five weighted-EM analytic controls. Independent 73840 native features,424580 child Gaussian scalars,8632 hidden values,4760 reference decisions,80 budgets and6903 candidate records; maximum embedding difference5.946e-14.14326 real streaming rows,metadata/permutation/missing controls.32-entry native view,18 reference changes and50 reused source/cloud links checked.
All14 eligible fits converge; two identity fallbacks. Primary retains exact132/147 positives and selects5/151 others versus E1128, precision96.35%/recall89.80%. Secondary127/147 and8/151; nontraining E11321/25 stars and2/18 others (91.30%/84%), primary19/25 and0/18.3100/3803 experimental entries; primary subset of3225 baseline. Full7041-row streams run about2133/2176 rows/s,215/217MiB child RSS;6.56MiB verified duplicate readbacks deleted. No downloads.
Next: Retain E112 and both frozen hierarchy alternatives; investigate independent cached astrometric/spectral context and noise-aware missing-view discrimination, then freeze before fresh broad-survey evidence. Keep original star/YSO/dust and103 molecular tiers, unresolved cold-source coverage and pending native upload permission. Full robust goal stays active.
Existing consumed development references, inherited class/target source selection and labelled interpretation, no fresh global precision. All3 HOPS extragalactic proxies still pass; nontraining whole quality STAR recovery21/38 rather than21/25. Native measurement covariance/censoring unmodeled. Hierarchy is deterministic posterior-product embedding, not a global likelihood; ranking and hidden RMSE do not universally improve. Cloud proximity is not membership/birth.