On kernel methods for covariates that are rankings