NeurIPS 2020
### Weston-Watkins Hinge Loss and Ordered Partitions

### Meta Review

This is a nice theoretical contribution to the field of multiclass classifiers, proving that a well known loss (the so called Weston-Watkins loss, defined as the sum of hinge losses over relative log-odd scores) is calibrated for a new discrete loss introduced in the paper. While the practical implications of the results are not direct (and in particular, how this result explains the good empirical performance of the WW loss), it provides new perspectives and solid foundations to analyze loss functions for multiclass problems. The paper is overall very clear, rigorous and polished.