T-COL: Generating Counterfactual Explanations for General User Preferences on Variable Machine Learning Systems