Testing the Generalization and Domain Stability of Compact Feature Representations for Photometric Supernova Classification

Photometric classification of supernovae increasingly requires models that are not only accurate within a single survey but also robust to changes in cadence, noise properties, filter coverage, and survey domain. We investigate the generalization and domain stability of a compact 16-feature representation for Type Ia supernova classification. The feature set consists of physically interpretable descriptors of brightness, color, variability, and temporal evolution. Using the Supernova Photometric Classification Challenge (SPCC) dataset as the reference domain, we confirm that a compact XGBoost classifier achieves strong within-survey performance, reaching an F1 score of 0.844 and a PR-AUC of 0.928 on a held-out test set. We then evaluate robustness under alternative classifiers, repeated resampling, feature perturbations, missing-band proxies, shortened temporal coverage, and cross-survey transfer to PLAsTiCC. The compact representation remains stable under resampling and moderate perturbations, but direct SPCC to PLAsTiCC transfer produces a substantial performance degradation. Class-conditional centroid analysis shows that the SPCC and PLAsTiCC Type Ia populations occupy different regions of the compact feature space. The cross-survey Ia centroid shift exceeds the Ia/non-Ia separation within either survey, indicating that the transfer gap is driven primarily by feature-space domain shift rather than classifier instability. These results show that compact physically interpretable features are robust within a survey and useful for diagnostic analysis, but are not automatically survey-invariant. Cross-survey deployment therefore requires feature harmonization, domain adaptation, or restriction to a smaller set of survey-stable features.