k-nearest-neighbors regression or classification

Mirrors the C# KNearestNeighbors class: predicts from the k training rows closest to each new observation – an inverse-squared-distance weighted average of their responses for regression, or their most common response for classification.

Usage

ml_knn(
  x,
  y,
  newdata,
  k,
  regression = TRUE,
  what = c("prediction", "neighbors", "bootstrap", "intervals"),
  seed = NULL,
  realizations = 1000,
  alpha = 0.1
)

Arguments

x

a numeric matrix or data frame of training predictors, one row per observation.

y

the training response, one value per row of x.

newdata

predictors to predict for, with the same number of columns as x.

k

the number of neighbors.

regression

TRUE (the default) for a weighted average; FALSE for the modal class.

what

which result to return: "prediction" (the default), "neighbors" (the indices of each new observation's neighbors), "bootstrap" (a prediction from one bootstrap resample of the training data), or "intervals" (bootstrapped prediction intervals).

seed

integer PRNG seed, used by "bootstrap" and "intervals"; NULL uses the clock.

realizations

the number of bootstrap resamples for "intervals". Default 1000.

alpha

the interval level for "intervals": 0.1 (the default) gives a 90% interval.

Value

for "prediction" and "bootstrap", a numeric vector; for "neighbors", a matrix of 0-BASED training-row indices with one row per row of newdata and k columns; for "intervals", a matrix with columns lower, median, upper, mean. Note "intervals" always reports percentiles, even for a classifier – upstream has no classification branch there.

Examples

x <- c(1, 2, 3, 4, 5, 6, 7, 8, 9, 10)
y <- c(10, 20, 30, 40, 50, 60, 70, 80, 90, 100)
ml_knn(x, y, newdata = 4.5, k = 2)
#> [1] 45