A Classifier-Based Approach to Multi-Class Anomaly Detection for Astronomical Transients

Automating real-time anomaly detection is essential for identifying raretransients in the era of large-scale astronomical surveys. Modern surveytelescopes are generating tens of thousands of alerts per night, and futuretelescopes, such as the Vera C. Rubin Observatory, are projected to increasethis number dramatically. Currently, most anomaly detection algorithms forastronomical transients rely either on hand-crafted features extracted fromlight curves or on features generated through unsupervised representationlearning, which are then coupled with standard machine learning anomalydetection algorithms. In this work, we introduce an alternative approach todetecting anomalies: using the penultimate layer of a neural network classifieras the latent space for anomaly detection. We then propose a novel method,named Multi-Class Isolation Forests (MCIF), which trains separate isolationforests for each class to derive an anomaly score for a light curve from thelatent space representation given by the classifier. This approachsignificantly outperforms a standard isolation forest. We also use a simplerinput method for real-time transient classifiers which circumvents the need forinterpolation in light curves and helps the neural network model inter-passbandrelationships and handle irregular sampling. Our anomaly detection pipelineidentifies rare classes including kilonovae, pair-instability supernovae, andintermediate luminosity transients shortly after trigger on simulated ZwickyTransient Facility light curves. Using a sample of our simulations that matchedthe population of anomalies expected in nature (54 anomalies and 12,040 commontransients), our method was able to discover 41±3 anomalies ( 75after following up the top 2000 ( 15shows that classifiers can be effectively repurposed for real-time anomalydetection.