Identifying new X-ray binary candidates in M31 using random forest classification

Abstract

Identifying X-ray binary (XRB) candidates in nearby galaxies requires distinguishing them from possible contaminants including foreground stars and background active galactic nuclei. This work investigates the use of supervised machine learning algorithms to identify high-probability X-ray binary candidates. Using a catalogue of 943 Chandra X-ray sources in the Andromeda galaxy, we trained and tested several classification algorithms using the X-ray properties of 163 sources with previously known types. Amongst the algorithms tested, we find that random forest classifiers give the best performance and work better in a binary classification (XRB/non-XRB) context compared to the use of multiple classes. Evaluating our method by comparing with classifications from visible-light and hard X-ray observations as part of the Panchromatic Hubble Andromeda Treasury, we find compatibility at the 90% level, although we caution that the number of source in common is rather small. The estimated probability that an object is an X-ray binary agrees well between the random forest binary and multiclass approaches and we find that the classifications with the highest confidence are in the X-ray binary class. The most discriminating X-ray bands for classification are the 1.7-2.8, 0.5-1.0, 2.0-4.0, and 2.0-7.0 keV photon flux ratios. Of the 780 unclassified sources in the Andromeda catalogue, we identify 16 new high-probability X-ray binary candidates and tabulate their properties for follow-up.

Probabilistic classification of X-ray sources applied to Swift-XRT and XMM-Newton catalogs

“…The precision of XRBs and their recall seem to be slightly higher and lower, respectively, but this depends on the weight chosen for the XRB class. The f 1 score of XRBs is also higher than a recent random forest classifier of X-ray sources in the literature (0.769 for Arnason et al 2020, however not based on the same source properties). We compared random forest performance to our classification on the test sample, both statistically and with manual inspection of a small subsample.…”

“…In the literature, no classification work has addressed the 2SXPS catalog so far. However, other similar catalogs, such as the XMM-Newton serendipitous source catalog (Webb et al 2020 in its latest version), are subject to classification works, for example Lin et al (2012) and Farrell et al (2015), respectively, using the "threshold rules" approach and a random forest algorithm, while Arnason et al (2020) tested several machine learning methods to classify Chandra X-ray sources in M31. Besides, a study preliminary to this work (Primorac 2015) shows that the Swift catalog has an interesting potential for X-ray source classification, with a classification based on selection criteria forming a decision tree with three classes: AGN, stars, and stellar-mass compact objects.…”

“…Machine learning algorithms have been shown to be efficient in automatically classifying X-ray sources, for example the classification of ROSAT bright sources or XMM-Newton variable sources (McGlynn et al 2004, Farrell et al 2015. For such applications, the random forest algorithm (Breiman 2001) is often preferred to other methods, such as neural networks or support vector machines, as it seems to show better results (e.g., Arnason et al 2020) and is theoretically better adapted to X-ray source classification: Indeed, it can be used even with a small number of features (here the source properties) or when they are not normalized, it does not require linearly separable classes, and it is useful even when the training sample is small or unbalanced. In practice, the random forest algorithm consists of the automatic generation of a large number of decision trees similar to the one described above, each trained on a subset of the training sample with a subset of the features (i.e., source properties).…”

A new automated tool for the spectral classification of OB stars

“…After optimizing the algorithms and training the models, we evaluated their performance on the test set. For the evaluation of the algorithms, we used the confusion matrix, which is widely used in problems of statistical classification (e.g., Mahabal et 2017; Arnason et al 2020). The confusion matrix has a table layout that visualizes the performance of a classifier using the test subset.…”

“…The former use known and labeled data as inputs, while the latter try to learn the various groups of the data from the data itself (for a full overview, see Baron 2019). Recently, thanks to improvements on software development and computational power, an increasing number of studies apply machine learning to a variety of problems in astronomy, taking advantage of its computational speed, as well as, its ability to handle large volumes of data (e.g., Mahabal et al 2008;Laurino et al 2011;Castro et al 2018;Pearson et al 2019;Clarke et al 2020;Arnason et al 2020).…”