Machine learning-based classification of type 1 active galactic nuclei

Quasars and Seyfert galaxies are two prototypes of type 1 active galactic nuclei (AGNs). The current view posits that they belong to a uniform population with continuous distributions of the bolometric luminosity and the contribution from host galaxy. Clarifying whether physical differences exist between them would be helpful in improving our understanding of the nature of accretion of supermassive black holes and the co-evolution with their host galaxies. Although the sample sizes and studies of quasars have been growing rapidly recent years, research on Seyfert galaxies continues to be limited. Methods used to select Seyfert galaxies in large surveys or to distinguish quasars and Seyfert galaxies in broad-line AGNs have also received scarce attention in the literature. We aim to provide a machine learning-based method to distinguish type 1 quasars and type 1 Seyfert galaxies based on photometric data, which could be applied to large all-sky surveys. The method can also be used to obtain relatively complete samples of Seyfert galaxy candidates when broad-line AGNs have been selected. A principle component analysis (PCA) was first applied to reduce the dimensionality and visualize the separation between quasars and Seyfert galaxies based on multidimensional data. Then we trained an supervised classifier to distinguish quasars and Seyfert galaxies with optical and mid-infrared photometric data. The results of PCA show that quasars and Seyfert galaxies are partially distinguishable. The supervised classifier based on a random forest (RF) algorithm achieves a good level of performance for type 1 quasars, with a F_1 score of 0.967. The performance for classifying Seyfert galaxies was worse, with a F_1 score of 0.791. Increasing the classification threshold can help derive a sample of Seyfert galaxies with a high level of purity. The classifiers were then applied to the cross-matched sample with DESI DR1. When compared with the results of spectroscopic classification, quasars identifications were also accompanied by a high level of precision and recall. For Seyfert galaxies identified by RF classifier, there are plenty of less luminous sources who are classified as quasars by spectroscopic method. If they are regarded as Seyfert galaxies, our classifier also performs well for predicting Seyfert galaxies. Both the median SEDs and the fraction of Seyfert galaxies derived by our classifier are similar to those in the training sample, while the fraction of Seyfert galaxies for spectroscopic classification are lower at all redshifts. These results indicate that our classifier has the capacity to recover the features of Seyfert galaxies-at least with respect to the ones that are similar to the Seyfert galaxies in the training sample.