Browse Teach Curate

Cost-and-Reward-Infused Metric Elicitation for Classifiers

Extends the Diagonal Linear Performance Metric Elicitation (DLPME) framework so an elicited multiclass classification metric also weighs bounded, classifier-specific costs and rewards (e.g. monetary cost, latency) that lie outside the confusion matrix. The practitioner's true metric is still recovered from query-efficient pairwise preferences between classifiers, now over a jointly (confusion-matrix, cost, reward) parameterized metric space.

Data Text

2501.00696

Metric elicitation is the problem of discovering which performance metric a practitioner actually cares about, by asking them to compare classifiers rather than state a formula. Existing methods, suc…