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

Abstract The classification of Tidal Disruption Events in large-scale photometric surveys is challenging because deterministic machine learning models produce overconfident misclassifications under varying data quality and low signal-to-noise conditions. Existing lightcurve-based approaches do not propagate measurement uncertainties properly into the classification itself, consequently generating brittle outputs. We present a host-agnostic, uncertainty-aware classification framework utilizing a Probabilistic Random Forest (PRF). Our pipeline extracts 11 characteristic features-including rise and decay timescales, blackbody temperature evolution, and pre-transient variability metrics–from nuclear transients in the ZTF alert stream. Relying exclusively on photometric lightcurve data without host galaxy information ensures effectiveness for the faint transient population expected from the Rubin Observatory. The PRF classifier, which treats feature measurements as distributions, was evaluated against XGBoost through a Leave-One-Out Cross-Validation strategy. The analysis demonstrates the PRF yields higher stability and robustness for ambiguous candidates compared to the predictions of XGBoost. The two classifiers occupy complementary regimes: XGBoost achieves higher recall in balanced and high-precision scenarios, where its rigid decision boundaries efficiently isolate TDE-like sources, while PRF rejects more false positives by penalizing sources with large feature uncertainties. Applying this validated framework to archival data identified 11 new candidate TDEs from the unclassified population and isolated 3 potential photometric TDEs within existing training labels previously misclassified as supernovae or active galactic nuclei. This work demonstrates that TDEs can be reliably identified using photometric lightcurve features, providing a host-agnostic framework. Uncertainty-aware, probabilistic classifiers are essential for the Rubin era to prevent the overconfident misclassifications inherent in deterministic models operating at low signal-to-noise.