Class-Conditional Label Noise in Astroparticle Physics

Class-conditional label noise characterizes classification tasks in which the training set labels are randomly flipped versions of the actual ground-truth. The analysis of telescope data in astroparticle physics poses this problem with a novel condition: one of the class-wise label flip probabilities is known while the other is not. We address this condition with an objective function for optimizing the decision thresholds of existing classifiers. Our experiments on several imbalanced data sets demonstrate that accounting for the known label flip probability substantially improves the learning outcome over existing methods for learning under class-conditional label noise. In astroparticle physics, our proposal achieves an improvement in predictive performance and a considerable reduction in computational requirements. These achievements are a direct result of our proposal's ability to learn from real telescope data, instead of relying on simulated data as is common practice in the field.