Adversarial random forests (ARFs) have recently been introduced as data synthetizer. Based on unsupervised RFs, ARFs rely on a recursive adversarial procedure in which trees progressively learn the structural properties of the data through alternating rounds of data generation and discrimination. The unsupervised classificatin task is achieved by introducing a synthetic response variable $y \in \{0,1\}$, where $y = 0$ denotes synthetic data and $y = 1$ denotes original data. During the discrimination phase, the objective is to distinguish original observations from synthetic ones obtained via marginal resampling of the original features in terminal nodes. The adversarial process stops when the prediction accuracy for $y$ falls below a predefined threshold.

While adversarial random forests (ARFs) have demonstrated strong performance in various low-dimensional settings, their behavior in high-dimensional contexts, such as omics data, remains less well understood. A key assumption of ARFs — that feature distributions are independent within terminal nodes — may be violated in high-dimensional settings, where a small subset of features can be highly predictive of the synthetic response variable $y$, while features not used for splitting may remain correlated within terminal nodes.

The high-dimensional adversarial random forests (HARFs) package addresses this limitation by identifying isolated regions of the feature space in which the independence assumption is more likely to hold. In addition to these isolated regions, the HARF procedure constructs a low-dimensional meta-space that captures the relationships among regions. Within each isolated region, separate ARFs are trained to better capture local data structures. To preserve dependencies between regions, meta-features representing region membership are used in the meta-space.

devtools::install_github("bips-hp/harf")

Using the built-in single_cell dataset, we illustrate how to train a HARF model.

data(single_cell)
harf_model <- h_arf(
 omx_data = single_cell[ , - which(colnames(single_cell)  == "cell_type")],
 cli_lab_data = data.frame(cell_type = single_cell$cell_type)
)

synth_single_cell <- h_forge(
  harf_obj = harf_model,
  n_synth = nrow(single_cell)
  )

  lung_single_cell <- h_forge(
      harf_obj = harf_model,
      n_synth = sum(single_cell$cell_type == "lung"),
      evidence = data.frame(cell_type = "lung")
     )

We refer to the package vignette for detailed examples and explanations.

vignette("harf")

• Fouodo, C. J. K., et al. (2026). High-dimensional adversarial random forests. Submission. Link don’t click.

• Watson, D. S., Blesch, K., Kapar, J. & Wright, M. N. (2023). Adversarial random forests for density estimation and generative modeling. In Proceedings of the 26th International Conference on Artificial Intelligence and Statistics. Link here.