Robust Learning via Conditional Prevalence Adjustment