dirf - Distances with Random Forest

dirf implements two tree-based kernel smoothers—DiNo (Distance with Nodes) and RanBu (Random Bushes)—that operate post hoc on a shallow random forest. Both methods reuse the leaf assignments of a fixed-depth forest (typically ~50 trees, max depth 5) and turn tree-based distances into similarity weights through Gaussian kernel smoothing weighting with bandwidth h. DiNo relies on the MRCA distance, while RanBu uses Breiman’s distance (1 − RF proximity).

The implementation builds on forests fitted with the ranger package: once a shallow forest is trained, predictions are obtained without retraining by weighting training responses according to the kernelized distances. This supports conditional expectation (distance-weighted mean) for point prediction and extends naturally to quantile regression via a weighted empirical CDF to estimate conditional quantiles.

Reference

Santos, T.M., Izbicki, R. and Esteves, L.G. (2025) DiNo and RanBu: Lightweight Predictions from Shallow Random Forests. arXiv preprint arXiv:2510.23624. Available at: https://arxiv.org/abs/2510.23624 (Accessed: 29 October 2025).

Installation

You can install the development version of dirf from Github using devtools:

# Install devtools if necessary
install.packages("devtools")

# Install dirf
devtools::install_github("tiagomendonca/dirf")

Dependencies: Ensure you have the following packages installed: ranger, data.table, purrr, tibble, RcppAlgos, ape, and stringr.

Examples

Below are examples of how to use the key functionalities of the dirf package.

Data preparation

library(dirf)

# Create a training dataset
set.seed(123)

df1 <- tibble::tibble(
  x1 = stats::rnorm(300, 0, 1),
  x2 = stats::rnorm(300, 0, 1),
  x3 = stats::rnorm(300, 0, 1),
  y = 2 * x1 + 1.5 * x2 + x3 + stats::rnorm(300, 0, 1)
)

df2 <- tibble::tibble(
  x1 = stats::rnorm(5, 0, 1),
  x2 = stats::rnorm(5, 0, 1),
  x3 = stats::rnorm(5, 0, 1),
  y = 2 * x1 + 1.5 * x2 + x3 + stats::rnorm(5, 0, 1)
)

Mean prediction

The following example demonstrates how to use DiNo and RanBu to predict the mean response:

# Train a Random Forest model
rf <- ranger::ranger(
  y ~ .,
  max.depth = 5,
  write.forest = TRUE,
  data = df1,
  num.trees = 10
)

# DiNo mean prediction
dino(
  rf = rf,
  df1 = df1,
  y = df1$y,
  df = df2,
  h = 0.1,
  type = "mean"
)

# A tibble: 5 × 1
  predicted
      <dbl>
1      3.14
2     -2.17
3      3.47
4     -4.69
5      3.85

# RanBu mean prediction
ranbu(
  rf = rf,
  df1 = df1,
  y = df1$y,
  df = df2,
  h = 0.1,
  type = "mean"
)

# A tibble: 5 × 1
  predicted
      <dbl>
1      3.12
2     -2.69
3      4.03
4     -4.58
5      4.11

Quantile regression

The next example shows how to predict quantiles using DiNo and RanBu:

# Train a Random Forest model with quantile regression
rf <- ranger::ranger(
  y ~ .,
  max.depth = 5,
  write.forest = TRUE,
  quantreg = TRUE,
  data = df1,
  num.trees = 10
)

# DiNo quantile prediction
dino(
  rf = rf,
  df1 = df1,
  y = df1$y,
  df = df2,
  h = 0.1,
  type = "quantile",
  quantiles = c(.025, .5, .975)
)

# A tibble: 5 × 3
  q_0.025 q_0.5 q_0.975
    <dbl> <dbl>   <dbl>
1   1.72   3.44    3.44
2  -4.73  -4.51   -2.85
3   3.44   3.62    4.12
4  -6.06  -4.45   -3.11
5  -0.179  4.24    4.24

# RanBu quantile prediction
ranbu(
  rf = rf,
  df1 = df1,
  y = df1$y,
  df = df2,
  h = 0.1,
  type = "quantile",
  quantiles = c(.025, .5, .975)
)

# A tibble: 5 × 3
  q_0.025 q_0.5 q_0.975
    <dbl> <dbl>   <dbl>
1    3.00  3.00    3.44
2   -3.35 -3.35   -3.35
3    4.12  4.12    4.12
4   -5.88 -4.45   -4.42
5    1.80  4.24    4.24