fk_regression... A faster implementation of univariate kernel regression

[Larsvanderlaan]

In case this is of interest, I stumbled upon an implementation of kernel regression in one dimension that offers significant speed improvements and scalability compared to npreg (and this is an understatement). The package is called FKSUM (an interesting name choice), and you can find a link to it here.

Below is the SuperLearner wrapper I have been using. I used the Nadaraya-Watson implementation, but `fk_regression' also supports local linear regression, which may be a better option.

SL.fastkernel <- function (Y, X, newX, family = gaussian(), obsWeights = rep(1,
                                                                               length(Y)), rangeThresh = 1e-07, ...)
{

  X <- as.matrix(X)
  newX <- as.matrix(newX)
  if(ncol(X) > 1) {
    stop("Univariate X kernel smooths only.")
  }
  fit <- FKSUM::fk_regression(X, Y, type = 'NW')
  pred <- predict(fit, xtest = newX)

  fit <- list(object = fit)
  class(fit) <- "SL.fastkernel"
  out <- list(pred = pred, fit = fit)
  return(out)
}

predict.SL.fastkernel <- function (object, newdata, ...)
{

  pred <- predict(object, xtest = as.matrix(newdata))

  return(pred)
}