Spectral risk-based studying utilizing unbounded losses
Authors: Matthew J. Holland, El Mehdi Haress
Summary: On this work, we take into account the setting of studying issues below a large class of spectral threat (or “L-risk”) capabilities, the place a Lipschitz-continuous spectral density is used to flexibly assign weight to excessive loss values. We receive extra threat ensures for a derivative-free studying process below unbounded heavy-tailed loss distributions, and suggest a computationally environment friendly implementation which empirically outperforms conventional threat minimizers when it comes to balancing spectral threat and misclassification error
