Currently the classification and regression algorithms in the neighbors module use a flat prior. They should be modified to compute a prior based on training data, and to optionally accept a user-defined prior.