Speaker
Description
We explore the use of machine learning for AGN selection from data
akin to the anticipated LSST survey (e.g., Savic et al. 2023). The
data is gathered from the AGN Data Challenge which comprises of two
primary sources: SDSS Stripe 82 and XMM-LSS. We also utilize data from
the DECam Local Volume Exploratoration Survey (DELVE). We employ
unsupervised clustering techniques, distinguishing our approach from
conventional methods that rely on labeled data by leveraging all of
the data. We further visualize the photometric, astrometric,
time-domain, and morphology data in real space in order to better
understand the latent space produced when passing the data through an
autoencoder. The comparative effectiveness of TensorFlow and PyTorch
in constructing autoencoders is explored -- aiming to provide a
foundational comparison of these frameworks to help guide us in
choosing the most suitable framework. Using the full parameter space,
a selection efficiency/completeness baseline is created using a Random
Forest. Our goal then is to improve upon this baseline by iterative
application of improvements to the algorithm(s) and feature space
choices.
