Spectral risk-based learning using unbounded losses
This work addresses robust risk assessment for machine learning models in scenarios with extreme losses, though it appears incremental as it builds on existing spectral risk frameworks.
The paper tackles learning problems under spectral risk functions with unbounded heavy-tailed losses, obtaining excess risk guarantees and proposing an efficient implementation that empirically outperforms traditional risk minimizers in balancing spectral risk and misclassification error.
In this work, we consider the setting of learning problems under a wide class of spectral risk (or "L-risk") functions, where a Lipschitz-continuous spectral density is used to flexibly assign weight to extreme loss values. We obtain excess risk guarantees for a derivative-free learning procedure under unbounded heavy-tailed loss distributions, and propose a computationally efficient implementation which empirically outperforms traditional risk minimizers in terms of balancing spectral risk and misclassification error.