HDBSCAN and Isolation Forest for Anomaly Detection in Unlabeled Light-Curve Data of Astronomical Transients

Author: William Zhang¹, Dr Daniel Muthukrishna²

Affiliation: 

¹International Academy High School of Michigan

²Massachusetts Institute of Technology (MIT) Kavli Institute for Astrophysics and Space Research

​

Abstract:​​​​

​

Astronomical light curves provide crucial insights into transient and variable celestial phenomena. The vast amount of data generated by missions like the Transiting Exoplanet Survey Satellite (TESS) necessitates advanced anomaly detection techniques to uncover rare astrophysical events. This study proposes a hybrid unsupervised machine learning framework combining a Variational Autoencoder (VAE) for dimensionality reduction, HDBSCAN for density-based clustering, and Isolation Forest for anomaly validation. The methodology was applied to 3,659 unlabeled TESS light curves, successfully identifying 76 anomalies consistently detected by both methods. Results demonstrate the robustness of integrating VAE-based feature extraction with complementary anomaly detection algorithms. The approach offers a scalable solution for discovering novel astrophysical phenomena without requiring labelled training data.

​

Keywords: anomaly detection, unsupervised learning, variational autoencoder, HDBSCAN, Isolation Forest​

​​

ISSN 3069-8200