|
Abstract:
|
This doctoral dissertation addresses the development and application
of advanced methods for analyzing the temporal variability of active galactic
nuclei (AGN) through the modeling of their optical light curves. The research
integrates unsupervised and generative learning techniques, by combining Self-
Organizing Maps (SOM) for data preprocessing and Conditional Neural
Processes (CNP) for light curve prediction.
For the first time in the study of AGN light curves, clustering via SOM has
been implemented for preprocessing, alongside the application of CNP for
modeling variability. This innovative approach facilitates a more effective
modeling of light curves characterized by uneven sampling and missing
observations.
The QNPy software package was developed and optimized for large-scale
parallel processing of extensive time series data. The proposed methodology was
validated using light curves from the All-Sky Automated Survey for SuperNovae
(ASAS-SN) and the SWIFT/BAT mission, covering a broad range of time scales
and variability.
The analysis prove that clustering light curves with SOM enhances the
performance of neural process, particularly for objects exhibiting simpler
variability patterns. The effects of SOM hyperparameters on clustering and
prediction performance were carefully examined. The models were validated
using loss function and mean squared error evaluations on real data.
The proposed methodology shows strong potential for scalable processing
of the large time-series data, anticipated in upcoming projects such as the Vera
C. Rubin Observatory’s Legacy Survey of Space and Time, enabling automated
classification, anomaly detection, and the extraction of scientifically significant
objects from catalogs containing hundreds of millions of sources. |