Fair Data Representation for Machine Learning at the Pareto Frontier