Interpreting the detection of anomalies in SDSS spectra

¹ Departamento de Física, Universidad de Antofagasta Avenida Angamos 601, Antofagasta, Chile
² Université Côte d’Azur, Observatoire de la Côte d’Azur, CNRS, Laboratoire Lagrange, 06000 Nice, France

^★ Corresponding authors: edgar.ortiz@ua.cl; mederic.boquien@oca.eu

Received: 10 July 2025
Accepted: 29 September 2025

Abstract

Context. The increasing use of machine-learning methods in astronomy introduces important questions about interpretability. The complexity and nonlinear nature of machine-learning methods means that it can be challenging to understand their decision-making process, especially when applied to the detection of anomalies. While these models can effectively identify unusual spectra, it remains a great challenge to interpret the physical nature of the flagged outliers.

Aims. We aim to bridge the gap between an anomaly detection and the physical understanding by combining deep learning with interpretable machine-learning (iML) techniques to identify and explain anomalous galaxy spectra from SDSS data.

Methods. We present a flexible framework that uses a variational autoencoder to compute multiple anomaly scores, including physically motivated variants of the mean-squared error. We adapted the iML LIME algorithm to spectroscopic data, systematically explored segmentation and perturbation strategies, and computed explanation weights that identified the features that are most likely to cause a detection. To uncover population-level trends, we normalized the LIME weights and applied clustering to 1% of the most strongly anomalous spectra.

Results. Our approach successfully separated instrumental artifacts from physically meaningful outliers and grouped anomalous spectra into astrophysically coherent categories. These include dusty metal-rich starbursts, chemically enriched H II regions with moderate excitation, and extreme emission-line galaxies with a low metallicity and hard ionizing spectra. The explanation weights agree with established emission-line diagnostics and enable a physically grounded taxonomy of spectroscopic anomalies.

Conclusions. Our work shows that an interpretable anomaly detection provides a scalable, transparent, and physically meaningful approach to exploring large spectroscopic datasets. Our framework opens the door for incorporating interpretability tools into quality control, follow-up targeting, and discovery pipelines in current and future surveys.