A derived label is a proxy label produced from other data rather than directly observed as the true target. Such labels can make training possible when direct labels are unavailable, but they may introduce measurement error, bias, or misleading correlations.