Jan Doersch

Gamma/Hadron Separation using Weakly Supervised Classification for MAGIC

As a system of two ground-based Imaging Air Cherenkov Telescopes, MAGIC is dedicated to measure the Cherenkov radiation of very-high-energy gamma rays. One important task in the analysis is the separation of the gamma-ray signal from a hadronic background. For this purpose, simulated gamma-ray events and measured data, which mostly originate from hadronically induced cascades, are used to apply a random forest classifier. Provided that every event is labelled either as signal or background, this method is called fully supervised classification. Another novel approach is the weakly supervised classification, requiring no signal or background label. This offers the possibility to use measured data exclusively and thus, the classification becomes independent of potentially inaccurate simulations. In this talk, the concept of weakly supervised learning will be introduced and first studies to compare it with the fully supervised method will be presented.