A Search for a More Robust Active Galactic Nucleus Classifier Using Machine Learning

We use various machine learning algorithms to explore the difference between galaxies with and without active galactic nuclei (AGNs). Because AGNs in low-mass galaxies ( log M * / M ⊙ ≲ 10 ) tend to be underpredicted by many selection methods, we emphasize those in this regime by including objects from the Active Dwarf Galaxy Database (ADGD) and sources therein. We collected a sample of roughly 6500 active and 53,000 inactive galaxies with IR, optical, and near-UV photometric data from ground and space-based observatories. We use a gradient-boosted trees classifier to determine the most effective balance of galaxies (low-mass versus massive; active versus inactive) for the accurate selection of AGNs. Next, we select the most important features from our datasets, reducing the number from 66 photometric and 39 spectroscopic features to three groups of fewer than 10 features each. Using a multilayer perceptron or gradient-boosted trees classifier, we were able to achieve selection accuracies near 98% and 60% for the total sample and low-mass sample, respectively. This simultaneous performance in the low-mass and full mass ranges indicate a step toward a classifier that is able to capture more low-mass AGNs without overfitting, sacrificing accuracy, or being computationally demanding.