Changing-State Quasars Detection using a Neural Differential Equation Approach

Abstract

In this work, we present a novel deep learning-based approach for the detection of changing-state quasars (CSQSOs), which are of critical importance in quasars (QSOs) research due to their substantial temporal variations in optical properties. Detecting and monitoring CSQSOs presents considerable challenges, including data sparsity and irregularity, complex and diverse behavior, large data volume, and limited labeled training data. To address these challenges, we propose an unsupervised anomaly detection based on neural differential equations that incorporates the irregularities in time series data and captures the physical characteristics of QSOs. By utilizing neural differential equations as a generative model to learn the characteristics of normal QSOs, and employing reconstruction error-based anomaly detection to identify CSQSOs, the proposed method mitigates issues related to the scarcity of labeled data and irregular sampling. The effectiveness of the proposed method is validated through extensive simulation studies using three types of datasets (mean shift, variance shift, and exponential shift) derived from real-world physical models. Our method outperforms existing methods in CSQSO detection by simultaneously modeling the irregular and stochastic nature of QSOs.