Classification of blazars based on data-driven approaches

Technical Summary: Classification of Blazars Based on Data-Driven Approaches

Problem Statement
Active Galactic Nuclei (AGNs), particularly blazars, exhibit distinctive variability in their optical light curves. However, robust classification of blazars (comprising BL Lac objects and Flat-Spectrum Radio Quasars, FSRQs) based solely on optical data is challenging. In the optical band, the highly variable nonthermal jet emission characteristic of blazars can be contaminated or damped by the steady light of the host galaxy and the moderately variable thermal emission from the accretion disk and Broad Line Region (BLR). Consequently, blazars can resemble normal AGNs, making optical variability alone an insufficient classifier without multiwavelength confirmation. While $\gamma$-ray data offers a cleaner jet signature, it is limited to the brightest sources and requires spectroscopic follow-up for definitive classification, which is resource-intensive. This work addresses the need for an alternative classification strategy that relies exclusively on optical light curves to distinguish blazars from non-blazar AGNs.

Methodology
The study utilizes optical light curve data from the Gaia DR3 (G-band, 2014–2017) and Pan-STARRS (composite light curves, 2010–2014) surveys. The dataset comprises:

• Known Sources: Approximately 2,500 blazars (from the Roma-BZCAT catalog) and a balanced set of ~2,500 non-blazar AGNs (from the Gaia catalog of variable AGNs).
• Unknown Sources: 21,733 AGN candidates lacking counterparts in other catalogs.

Feature extraction was performed using the FATS (Feature Analysis for Time Series) library to generate standard statistical descriptors (e.g., amplitude, autocorrelation, flux percentiles) and custom algorithms based on literature methods (e.g., Damped Random Walk models). A total of ~70 features were initially extracted and normalized.

The core methodology employs two machine learning approaches:

1. Supervised Classification (LightGBM): A Light Gradient-Boosting Machine model was trained on the labeled "Known" dataset using a five-fold cross-validation strategy. Hyperparameters ($learning\_rate$, $n\_estimators$, $colsample\_bytree$) were optimized via GridSearchCV.
2. Semi-Supervised Learning (Self-Training Classifier - STC): To classify the "Unknown" objects, a self-training loop was implemented. An initial LightGBM model trained on known data iteratively assigned pseudo-labels to unlabeled objects with high confidence (threshold > 0.95), expanding the training set until convergence.

Feature Selection
To address the high dimensionality of the feature space, the BoostBoruta algorithm (an extension of Boruta using gradient boosting) was applied. This method identifies relevant features by comparing original features against randomized "shadow" features. SHAP (SHapley Additive exPlanations) values were integrated to enhance feature selection and interpretability, quantifying individual feature contributions to model predictions.

Key Results

• Supervised Performance: The LightGBM model achieved an accuracy of 86% (±0.012) on the known dataset. Precision, recall, and F1 scores for both blazar and non-blazar classes exceeded 80–85%.
• Feature Reduction: BoostBoruta successfully reduced the feature space from ~70 to 13 key features (including Mean, Eta_e, DRW_tau, CAR_sigma, etc.) while maintaining performance comparable to the full feature set (Accuracy: 0.856 ± 0.012).
• Semi-Supervised Performance: The Self-Training Classifier achieved an accuracy of 85% (±0.013) on the expanded dataset including the 21,733 unknown objects. The performance metrics remained consistent with the supervised baseline, validating the reliability of the pseudo-labeling approach.
• Feature Importance: The three most critical features identified across all experiments were Mean (average magnitude), Eta_e (ratio of mean squared successive magnitude differences to variance), and DRW_tau (relaxation time from a Damped Random Walk model). These features reflect the physical dichotomy between the rapid, intense variability of blazars and the smoother evolution of disk-dominated AGNs.

Significance and Claims
The paper claims that its primary contribution is demonstrating the feasibility of classifying blazars versus non-blazar AGNs using only optical time-series data, without relying on multiwavelength information (radio, X-ray, $\gamma$-ray) or spectroscopic follow-up. The authors emphasize that this approach leverages the abundance and accessibility of data from large-scale surveys like Gaia, Pan-STARRS, and the upcoming Rubin-LSST.

The study highlights that despite the physical overlap in optical variability signatures between blazars and other AGNs, data-driven methods can achieve robust discrimination. The successful application of self-training to a large pool of unlabeled candidates suggests that this methodology can scale to future surveys, potentially identifying tens of thousands of new blazar candidates. The work establishes a foundation for efficient population-level studies of AGNs using optical variability alone, addressing the limitations of current classification bottlenecks.