On loss functions and regret bounds for multi-category classification

TL;DR

This paper introduces new loss functions and regret bounds for multi-class classification, improving theoretical understanding and providing tighter, simpler convex losses with strong regret guarantees.

Contribution

It develops novel inverse mappings for constructing proper scoring rules and hinge-like losses, revealing relationships among existing losses and establishing general regret bounds.

Findings

New inverse mappings from generalized entropy to losses

Identification of new multi-class proper scoring rules

Derivation of tighter, simpler hinge-like convex losses

Abstract

We develop new approaches in multi-class settings for constructing proper scoring rules and hinge-like losses and establishing corresponding regret bounds with respect to the zero-one or cost-weighted classification loss. Our construction of losses involves deriving new inverse mappings from a concave generalized entropy to a loss through the use of a convex dissimilarity function related to the multi-distribution $f$-divergence. Moreover, we identify new classes of multi-class proper scoring rules, which also recover and reveal interesting relationships between various composite losses currently in use. We establish new classification regret bounds in general for multi-class proper scoring rules by exploiting the Bregman divergences of the associated generalized entropies, and, as applications, provide simple meaningful regret bounds for two specific classes of proper scoring rules. Finally, we derive new hinge-like convex losses, which are tighter convex extensions than related hinge-like losses and geometrically simpler with fewer non-differentiable edges, while achieving similar regret bounds. We also establish a general classification regret bound for all losses which induce the same generalized entropy as the zero-one loss.