A two-stage approach where a model first learns representations, then retrains just the final classification layer on balanced data.