Semi-Supervised Empirical Risk Minimization: When can unlabeled data improve prediction