Emphasizing feature inter-class separability for improving highly imbalanced overlapped data classification.