DirichletRF

DirichletRF Logo

An R package implementing a parallel Dirichlet Random Forest for modeling compositional data, built on a Dirichlet log-likelihood splitting criterion and OpenMP-accelerated tree construction.

📦 Installation

⭐ Recommended: Install from GitHub (latest version)

# devtools::install_github("Xaleed/DirichletRF")
install.packages("DirichletRF", repos = c('https://xaleed.r-universe.dev', 'https://cloud.r-project.org'))

The GitHub version includes several features not yet available on CRAN:

Out-of-bag (OOB) predictions and error estimation
Distributional predictions and proximity weights
Feature importance (gain, split count, permutation)

Install from CRAN (limited features)

install.packages("DirichletRF")

⚠️ The CRAN version does not include OOB estimation, distributional mode, or importance measures. Use the GitHub version for full functionality.

Windows binary (no Rtools required)

install.packages(
  "https://github.com/Xaleed/DirichletRF/releases/download/v0.1.0/DirichletRF_0.1.0.zip",
  repos = NULL, type = "win.binary"
)

🚀 Quick Start

library(DirichletRF)

set.seed(42)
n <- 200; p <- 5
X <- matrix(rnorm(n * p), n, p)
colnames(X) <- paste0("X", 1:p)

# X1 and X2 are informative
alpha_mat <- cbind(2 + 3 * (X[,1] > 0), 3 + 3 * (X[,2] > 0), rep(4, n))
G <- matrix(rgamma(n * 3, shape = as.vector(t(alpha_mat))), n, 3, byrow = TRUE)
Y <- G / rowSums(G)

forest <- DirichletRF(X, Y, num.trees = 100)
print(forest)

🔧 Features & Examples

Prediction

Xtest <- matrix(rnorm(10 * p), 10, p)
colnames(Xtest) <- paste0("X", 1:p)

pred <- predict(forest, Xtest)
pred$mean_predictions                              # mean-based predictions
pred$alpha_predictions / rowSums(pred$alpha_predictions)  # parameter-based

OOB Estimation

Requires replace = TRUE or sample.fraction < 1. Not available in the CRAN version.

# Bootstrap OOB
forest_oob <- DirichletRF(X, Y, num.trees = 100,
                           replace = TRUE, compute.oob = TRUE)
forest_oob$oob$mse
forest_oob$oob$predictions   # n x k matrix, NA where never OOB

# Subsampling OOB
forest_sub <- DirichletRF(X, Y, num.trees = 100,
                           replace = FALSE, sample.fraction = 0.632,
                           compute.oob = TRUE)
forest_sub$oob$mse

Feature Importance

Not available in the CRAN version.

# Impurity-based: gain and split count
importance(forest)

# Permutation-based (requires compute.oob = TRUE)
permutation_importance(forest_oob, X, loss = "aitchison",
                       num.permutations = 10, seed = 42L)

Both return a data frame sorted by importance descending. permutation_importance() supports "aitchison" (default), "mse", and "kl" as loss functions.

Distributional Mode & Proximity Weights

Stores leaf indices so the forest can return a weighted distribution over training observations for any test point. Not available in the CRAN version.

forest_dist <- DirichletRF(X, Y, num.trees = 100, distributional = TRUE)

# n_test x n_train weight matrix; rows sum to 1
W <- predict_weights(forest_dist, Xtest)

# Weighted conditional mean
Y_hat <- W %*% Y

# Draw samples from the conditional distribution P(Y | x_new)
draws <- sample_conditional(forest_dist, x_new = Xtest[1, ], size = 200L)
colMeans(draws)   # estimated conditional mean

OOB Proximity Matrix

Available when both distributional = TRUE and compute.oob = TRUE.

forest_full <- DirichletRF(X, Y, num.trees = 100,
                            distributional = TRUE,
                            replace = TRUE, compute.oob = TRUE)

W_oob <- forest_full$oob$weights   # n x n, generally asymmetric
W_sym <- (W_oob + t(W_oob)) / 2   # symmetrise if needed

⚙️ Main Parameters

Parameter	Default	Description
`num.trees`	`100`	Number of trees
`max.depth`	`10`	Maximum tree depth
`min.node.size`	`5`	Minimum leaf size
`mtry`	`-1` (= √p)	Candidate features per split
`est.method`	`"mom"`	`"mom"` or `"mle"`
`num.cores`	`-1` (all−1)	OpenMP threads
`replace`	`FALSE`	Bootstrap or subsampling
`sample.fraction`	`1.0`	Fraction of data per tree
`compute.oob`	`FALSE`	Compute OOB predictions
`distributional`	`FALSE`	Store leaf indices for weight-based predictions

Note: X must be numeric. Use one-hot encoding for categorical covariates. Rows of Y must sum to 1.

📚 Reference

Masoumifard, K., van der Westhuizen, S., & Gardner-Lubbe, S. (2026). Dirichlet random forest for predicting compositional data. In A. Bekker et al. (Eds.), Environmental Modelling with Contemporary Statistics. Chapman & Hall/CRC. ISBN: 9781032903910.

📄 License

GPL-3