How to Find Variable Active Galactic Nuclei with Machine Learning

TL;DR

This paper demonstrates an unsupervised machine learning approach using self-organizing maps to identify variable active galactic nuclei in large datasets, achieving high purity and completeness, and enabling analysis of physical correlations.

Contribution

The study introduces a novel unsupervised SOM-based method for detecting variable AGN, maintaining domain knowledge and systematics control, with performance comparable to supervised neural networks.

Findings

Achieved 86% purity and 66% completeness in identifying variable AGN.

Validated the method on simulated and real WISE data.

Applicable to various time-sampled light curves beyond AGN.

Abstract

Machine-learning (ML) algorithms will play a crucial role in studying the large datasets delivered by new facilities over the next decade and beyond. Here, we investigate the capabilities and limits of such methods in finding galaxies with brightness-variable active galactic nuclei (AGN). Specifically, we focus on an unsupervised method based on self-organizing maps (SOM) that we apply to a set of nonparametric variability estimators. This technique allows us to maintain domain knowledge and systematics control while using all the advantages of ML. Using simulated light curves that match the noise properties of observations, we verify the potential of this algorithm in identifying variable light curves. We then apply our method to a sample of ~8300 WISE color-selected AGN candidates in Stripe 82, in which we have identified variable light curves by visual inspection. We find that with ML we can identify these variable classified AGN with a purity of 86% and a completeness of 66%, a performance that is comparable to that of more commonly used supervised deep-learning neural networks. The advantage of the SOM framework is that it enables not only a robust identification of variable light curves in a given dataset, but it is also a tool to investigate correlations between physical parameters in multi-dimensional space - such as the link between AGN variability and the properties of their host galaxies. Finally, we note that our method can be applied to any time-sampled light curve (e.g., supernovae, exoplanets, pulsars, and other transient events).