Balanced Self-Paced Learning for AUC Maximization