harf: High-Dimensional Adversarial Random Forests for Omics Data

Introduction

Adversarial random forests (ARFs) have recently been introduced as a well-performing generative method for low-dimensional datasets. 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 classification 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. Synthetic data are created during the the adversarial game by marginal resampling of the original features in terminal nodes. The adversarial process stops when the prediction accuracy for $y$ falls below a predefined threshold, $0.5$, for example.

While ARFs have demonstrated strong performance in various low-dimensional settings, their behavior in high-dimensional contexts, such as omics data, remains non investigated. The key assumption of ARFs — that feature distributions are independent within terminal nodes — may be violated in high-dimensional settings. For example, if a small subset of features is highly predictive of the synthetic response variable $y$, the remaining features not used to reach the terminal node may still be correlated, not allowing for marginal resampling. A typical resulting behaviour of ARF in such a situation is that the algorithm terminates without converging.

The high-dimensional adversarial random forests (HARFs) package addresses the convergence issue of ARF by identifying isolated regions of the feature space in which the independence assumption is more likely to hold in terminal nodes. In addition, 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 synthesize observations, the HARF procedure first samples a region from the meta-space and then condition the resampling in the other isolated region by synthesized meta-observation. HARF offers a flexible framework for unconditional and conditional data generation for clustering downstream analysis, as well as for prediction tasks. The package includes a built-in single-cell RNA-seq datasets and TCGA-KICH dataset to illustrate its usage. We refer to the package vignette for detailed examples and explanations.

Package installation

devtools::install_github("bips-hb/harf", build_vignettes  = TRUE)

vignette("harf")

Vignette on browser

browseVignettes("harf")

References

• Fouodo, C. J. K., Kapar, J. & Wright, M. N. (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.