🔭 astrophysics

Uncertainty-Aware Tidal Disruption Event Classification : A Host-Agnostic Probabilistic Random Forest Approach

This paper introduces a host-agnostic, uncertainty-aware Probabilistic Random Forest framework that leverages photometric lightcurve features to robustly classify Tidal Disruption Events by treating measurement uncertainties as distributions, thereby reducing overconfident misclassifications common in deterministic models and identifying new candidates for future Rubin Observatory surveys.

Original authors: Vysakh Anilkumar, Sjoert van Velzen, Marek Kowalski, Simeon Reusch

Published 2026-07-31

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Vysakh Anilkumar, Sjoert van Velzen, Marek Kowalski, Simeon Reusch

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the night sky as a giant, bustling city of stars. Most of these stars are quiet and predictable, like neighbors who keep to their routine. But sometimes, something dramatic happens: a supermassive black hole, the invisible "monster" lurking at the center of a galaxy, decides to eat a star that wandered too close. When this happens, the star gets stretched into a long strand of spaghetti and then devoured, creating a massive, glowing flare of light. Astronomers call this a Tidal Disruption Event, or TDE. It's like a cosmic fireworks display that can last for months or even years.

The problem is that the universe is huge, and these events are rare and often faint. To find them, astronomers use giant cameras on telescopes that take pictures of the sky over and over again, creating a massive stream of data. The challenge is sorting through millions of these "flashes" to find the real TDEs. It's like trying to find a specific type of firework in a pile of garbage, sparklers, and lightning bugs. For a long time, scientists relied on looking at the host galaxy (the neighborhood where the flash happened) or getting a detailed "spectrum" (a chemical fingerprint) of the light to confirm what they found. But with new, super-powerful telescopes coming online, we are going to find so many of these events that we won't have time to check the neighborhood or get a fingerprint for every single one. We need a way to identify them just by looking at how the light changes over time, and we need to be smart about how much we trust that light, especially when the signal is weak and fuzzy.

This paper introduces a new, clever way to do exactly that. The authors built a computer program designed to spot TDEs using only the light curves (the story of how the brightness changes over time) without needing to know anything about the galaxy hosting the event. They call their method "host-agnostic," which is a fancy way of saying "it doesn't care where the event happened."

To solve the problem of fuzzy, low-quality data, they didn't just build a standard computer brain; they built one that knows when it's unsure. They compared two different types of machine learning models. The first, called XGBoost, is like a very strict, confident detective. It looks at the clues and makes a sharp, binary decision: "This is a TDE" or "This is not." It's great when the clues are clear, but if the clues are blurry, it might still shout "Guilty!" with 100% confidence, even if it's wrong. The second model, which the authors developed and call a "Probabilistic Random Forest" (PRF), is more like a cautious detective who keeps a notebook of uncertainties. Instead of treating every measurement as a perfect fact, it treats them as a range of possibilities. If a piece of data is fuzzy, this detective says, "Well, it could be a TDE, but I'm not entirely sure because the data is messy."

The researchers tested these models on data from the Zwicky Transient Facility (ZTF), a telescope that has already found over 100 TDEs. They found that while the strict detective (XGBoost) is good at catching clear-cut cases, the cautious detective (PRF) is much better at handling the messy, faint, and ambiguous ones. The PRF model is more stable; it doesn't flip-flop wildly when the data changes slightly, and it is much better at saying "I don't know" or "This is probably a false alarm" when the signal is weak. In fact, the PRF model rejected about 35% more fake candidates (false positives) than the strict model did when looking for as many real events as possible.

Using this new, uncertainty-aware system, the team went back through old data and found 11 new TDE candidates that had been missed or misclassified as something else, like a supernova or an active black hole. They also spotted three potential TDEs that had been previously labeled as supernovae or active galaxies, suggesting those labels might have been wrong. The paper concludes that as we move toward the era of even bigger telescopes (like the Rubin Observatory), which will find millions of faint events, we need these "uncertainty-aware" tools. We can't afford to have computers confidently misclassifying faint, fuzzy signals; we need them to admit when the data is too noisy to be sure, ensuring we don't waste time chasing ghosts in the night sky.

Technical Summary: Uncertainty-Aware Tidal Disruption Event Classification

Problem Statement
The classification of Tidal Disruption Events (TDEs) in large-scale photometric surveys faces two critical bottlenecks as astronomy transitions toward the Vera C. Rubin Observatory era. First, existing machine learning models often rely heavily on host galaxy information (e.g., nuclear offsets, post-starburst indicators). This introduces bias and limits applicability to faint transients where host data (such as WISE or spectroscopic follow-up) is unavailable. Second, current deterministic classifiers (e.g., standard Random Forests or XGBoost) treat feature measurements as exact point estimates. This approach fails to propagate measurement uncertainties, leading to brittle, overconfident misclassifications when processing the sparse, low signal-to-noise (SNR) lightcurves expected from fainter candidates.

Methodology
The authors propose a host-agnostic, uncertainty-aware classification framework utilizing a Probabilistic Random Forest (PRF). The methodology consists of three primary stages:

Data Preparation and Feature Engineering:
- Dataset: The study utilizes a nuclear transient dataset from the Zwicky Transient Facility (ZTF) containing 12,251 sources. Labels include 59 spectroscopically confirmed TDEs, 1,042 Supernovae (SNe), and 5,113 Active Galactic Nuclei (AGN).
- Lightcurve Modeling: A multi-band lightcurve model is fitted to $g$ and $r$ band data. The model combines a Gaussian rise and exponential decay (transient component) with a sigmoid function modeling a late-time flux plateau. A blackbody scaling term accounts for temperature evolution.
- Parameter Inference: Markov Chain Monte Carlo (MCMC) sampling is used to derive posterior distributions for model parameters (e.g., rise time $t_{rise}$ , decay time $t_{decay}$ , peak temperature $T_{peak}$ , plateau strength). This step explicitly captures measurement uncertainties.
- Feature Extraction: 11 characteristic features are extracted, including timescales, temperature evolution metrics, plateau signal-to-noise ratios, and pre-transient fractional RMS variability. Crucially, the pipeline outputs both the feature values and their associated uncertainties.
Classification Framework:
- Probabilistic Random Forest (PRF): The primary classifier treats feature inputs as probability density functions rather than point estimates. At each decision tree node, the PRF calculates the probability that an object satisfies a split condition given its measurement uncertainty, propagating the object to both child branches with weighted probabilities. This naturally handles missing values (by treating them as infinite uncertainty) and heteroscedasticity.
- Baseline (XGBoost): An Extreme Gradient Boosting (XGBoost) classifier is trained as a deterministic baseline. It treats features as exact point estimates, ignoring measurement errors.
- Training Strategy: To address severe class imbalance (TDEs $\approx$ 1% of the sample), an uncertainty-aware extension of SMOTE (Synthetic Minority Over-sampling TEchnique) is applied to the training set. Performance is evaluated using a modified Leave-One-Out Cross-Validation (LOOCV) strategy with 250 random repetitions to assess stability.

Key Results

Stability and Robustness: The PRF demonstrates significantly higher stability than XGBoost, particularly for ambiguous candidates. While XGBoost predictions for uncertain sources fluctuate erratically across training iterations (often jumping between confident TDE and non-TDE classifications), PRF produces consistent, broader probability distributions that honestly reflect data quality.
Performance Trade-offs:
- In high-precision scenarios (requiring $\approx$ 85% purity), XGBoost achieves slightly higher recall ( $\approx$ 44% vs. 34% for PRF) by enforcing rigid decision boundaries.
- In high-recall scenarios (requiring $\approx$ 95% completeness), PRF is substantially more efficient, rejecting $\approx$ 35% more false positives (143 FPs vs. 220 FPs) than XGBoost.
Feature Importance: SHAP analysis reveals that PRF prioritizes stable, low-variance metrics (e.g., pre-transient variability) and down-weights features with high uncertainty (e.g., peak temperature derived from multi-band fitting). Conversely, XGBoost relies heavily on point estimates of physical parameters like $T_{peak}$ , making it blind to the uncertainty inherent in those measurements.
Application to Real Data:
- Validation: When tested on 12 TDEs discovered after the training cutoff, both models performed well, though PRF provided more conservative probability estimates for ambiguous cases.
- Discovery: Applying the ensemble to archival ZTF data identified 11 new TDE candidates from previously unclassified sources. Additionally, the framework flagged 3 potential misclassifications within existing labels: one supernova (SN2019meh) and two AGN (AT2020actcj and AT2019gtmk) that exhibit TDE-like lightcurve properties.
Low-SNR Performance: Simulated signal degradation tests show that PRF probability scores degrade monotonically as photometric quality decreases, whereas XGBoost exhibits an abrupt collapse in confidence, highlighting PRF's suitability for the faint end of the survey.

Significance and Claims
The paper argues that uncertainty-aware, probabilistic classifiers are essential for the Rubin Observatory era, where the majority of TDE candidates will be too faint for spectroscopic follow-up. The authors claim that:

Host-Agnostic Viability: TDEs can be reliably identified using photometric lightcurve features alone, without relying on host galaxy properties that may be unavailable for faint sources.
Necessity of Uncertainty Propagation: Deterministic models produce overconfident misclassifications in low-SNR regimes. By propagating measurement uncertainty through the classification process, PRF prevents the "brittle" behavior seen in standard models.
Complementary Architectures: The work does not position PRF as a strict replacement for deterministic models but rather as a complementary tool. XGBoost remains effective for well-constrained, high-SNR data, while PRF is superior for handling the sparse, noisy data expected from future deep surveys.

The authors conclude that this framework provides a necessary foundation for reliable TDE identification in the coming decade of time-domain astronomy, though they note future work is needed to adapt feature extraction for the irregular sampling cadence of the Rubin Observatory.

More like this