Training Classifiers under Covariate Shift by Constructing the Maximum Consistent Distribution Subset