Gradient Weights help Nonparametric Regressors

We show that weighting each coordinate i with the estimated norm of the ith derivative of f is an efficient way to significantly improve the performance of distance-based regressors, e.g.