Leveraging Ensemble Diversity for Robust Self-Training in the Presence of Sample Selection Bias