Interpreting automatic AGN classifiers with saliency maps

The classification of the optical spectra of active galactic nuclei (AGN) into different types is well founded on AGN physics, but it involves some degree of human oversight and cannot be reliably scaled to large data sets. Machine learning (ML) tackles such a classification problem in a fast and reproducible way, but is often perceived as a black box. However, ML interpretability and explainability are active research areas in computer science, increasingly providing us with tools to alleviate this issue. We applied ML interpretability tools to a classifier trained to predict AGN type from spectra, to demonstrate the use of such tools in this context. We trained a support-vector machine on 3346 high-quality, low redshift AGN spectra from SDSS DR15 with an existing reliable classification as type 1, type 2, or intermediate type. On a selection of test-set spectra, we computed the gradient of the predicted class probability and we built saliency maps. We also visualized the high-dimensional space of AGN spectra using t-distributed stochastic neighbor embedding (t-SNE), showing where the spectra for which we computed a saliency map are located. Regions that affect the predicted AGN type often coincide with physically relevant features, such as spectral lines. t-SNE visualization shows good separability of type 1 and type 2 spectra, while intermediate-type spectra either lie in-between as expected or appear mixed with type 2 spectra. Saliency maps show why a given AGN type was predicted by our classifier, resulting in a physical interpretation in terms of regions of the spectrum that affected its decision, making it no longer a black box. These regions coincide with those used by human experts such as relevant spectral lines, and are even used in a similar way, with the classifier e.g. effectively measuring the width of a line by weighing its center and its tails oppositely.