Combining Accuracy and Prior Sensitivity for Classifier Design Under Prior Uncertainty