Paper proposes adaptive margin loss to curb center-class bias in ordinal classification

A new arXiv preprint introduces Adaptive Margin Ordinal Loss, a training objective aimed at a problem the authors call center-class hedging. They argue that standard cross-entropy encourages networks to favor middle categories on ordinal tasks, since that choice lowers expected error, and their method adds class-dependent margins to discourage this. The work targets research on ordinal classification, where labels have a meaningful order, such as severity ratings or age brackets.