From Principle to Practice: Vertical Data Minimization for Machine Learning