Optimizing Supernova Classification: A Metric-Aware Approach with PR-AUC and F1-score

Photometric classification of Type Ia supernovae (SNe Ia) is essential for cosmological studies but remains challenging due to severe class imbalance and observational noise. While recent studies have explored deep learning models, these approaches are often resource-intensive and complex to interpret.This study aims to develop a computationally efficient and interpretable classification framework that maintains high performance on imbalanced supernova datasets. Rather than proposing a new metric, we emphasize the proper use of established ones—particularly PR-AUC and F1-score as more informative alternatives to ROC-AUC and SPCC-F1 in imbalanced settings.We use an XGBoost ensemble model, optimized through Bayesian hyperparameter tuning. The dataset, sourced from the Supernova Photometric Classification Challenge (SPCC), consists of 21,318 light curves with a class imbalance ratio of 3.19 (non-Ia to Ia). Feature engineering and threshold tuning were applied to improve classification precision and recall. Evaluation was based on PR-AUC, F1-score, and ROC-AUC for comparative benchmarking.Our model achieved a PR-AUC of 0.993, an F1-score of 0.923, and a ROC-AUC of 0.976, matching or exceeding deep learning-based models on precision-recall metrics, while trading off slightly on overall accuracy. Despite lower accuracy than some complex architectures, our method yields superior balance between false positives and false negatives.This study reinforces that ensemble learning models like XGBoost, when carefully optimized and evaluated with appropriate metrics, offer a robust, reproducible, and computationally lightweight solution for supernova classification. The results suggest that simpler models can be competitive alternatives to deep learning, particularly in large-scale surveys such as LSST where efficiency and transparency are critical.