We introduce the Fair Minimum Labeling problem for designing temporally efficient and fair activation plans, prove tight hardness bounds, and present approximation algorithms with strong empirical results on fair multi-source learning.