Improving Interaction Modelling and Interpretability in Random Forests

Implementation of the unity forest (UFO) framework (Hornung & Hapfelmeier, 2026, ). UFOs are a random forest variant designed to better take covariates with purely interaction-based effects into account, including interactions for which none of the involved covariates exhibits a marginal effect. While this framework tends to improve discrimination and predictive accuracy compared to standard random forests, it also facilitates the identification and interpretation of (marginal or interactive) effects: In addition to the UFO algorithm for tree construction, the package includes the unity variable importance measure (unity VIM), which quantifies covariate effects under the conditions in which they are strongest - either marginally or within subgroups defined by interactions - as well as covariate-representative tree roots (CRTRs) that provide interpretable visualizations of these conditions. Categorical and continuous outcomes are supported. This package is a fork of the R package 'ranger' (main author: Marvin N. Wright), which implements random forests using an efficient C++ backend.

Conformal Random Forests for Response Surface Emulation

Fits emulators, also known as surrogates or response surfaces, using conformal inference with random forests. The conformal calibration is performed using out-of-bag samples from the forest, eliminating the need for a separate hold-out set. The method is based on Johansson et al. (2014 ).

Multiple Imputation Using MICE and Random Forest

Functions to impute using random forest under full conditional specifications (multivariate imputation by chained equations). The methods are described in Shah and others (2014) .

Oblique Decision Random Forest for Classification and Regression

The oblique decision tree (ODT) uses linear combinations of predictors as partitioning variables in a decision tree. Oblique Decision Random Forest (ODRF) is an ensemble of multiple ODTs generated by feature bagging. Oblique Decision Boosting Tree (ODBT) applies feature bagging during the training process of ODT-based boosting trees to ensemble multiple boosting trees. All three methods can be used for classification and regression, and ODT and ODRF serve as supplements to the classical CART of Breiman (1984) and Random Forest of Breiman (2001) respectively.

Fast Serializable Random Forests Based on 'ranger'

An updated implementation of R package 'ranger' by Wright et al, (2017) for training and predicting from random forests, particularly suited to high-dimensional data, and for embedding in 'Multiple Imputation by Chained Equations' (MICE) by van Buuren (2007) . Ensembles of classification and regression trees are currently supported. Sparse data of class 'dgCMatrix' (R package 'Matrix') can be directly analyzed. Conventional bagged predictions are available alongside an efficient prediction for MICE via the algorithm proposed by Doove et al (2014) . Trained forests can be written to and read from storage. Survival and probability forests are not supported in the update, nor is data of class 'gwaa.data' (R package 'GenABEL'); use the original 'ranger' package for these analyses.

Approximate Bayesian Computation via Random Forests

Performs Approximate Bayesian Computation (ABC) model choice and parameter inference via random forests. Pudlo P., Marin J.-M., Estoup A., Cornuet J.-M., Gautier M. and Robert C. P. (2016) . Raynal L., Marin J.-M., Pudlo P., Ribatet M., Robert C. P. and Estoup A. (2019) .

Random Forest-Based Multistate Survival Analysis

Fits transition-specific cause-specific random survival forests on a clock-reset duration scale for acyclic, non-recurrent multistate processes. Entry-conditioned state-occupation probabilities are assembled from predicted cumulative hazards by semi-Markov entry-mass and sojourn convolution on a validated regular grid. The one-row-per-subject interface supports one common initial state, one recorded entry per state, baseline time-fixed covariates, competing exits, and independent right censoring. Left truncation, recurrent visits, directed cycles, time-dependent covariates, and ongoing-sojourn dynamic prediction are not supported. The package also provides calendar-time Aalen-Johansen point estimates as a covariate-free descriptive baseline, transition-specific permutation importance, genuine ranger edge OOB concordance, and patient-level cross-validated IPCW state-probability scoring. Methods are described in Ishwaran et al. (2008) for random survival forests, Putter et al. (2007) for multistate competing risks decomposition, and Aalen and Johansen (1978) < https://www.jstor.org/stable/4615704> for the nonparametric estimator.

Stepwise Predictive Variable Selection for Random Forest

An introduction to several novel predictive variable selection methods for random forest. They are based on various variable importance methods (i.e., averaged variable importance (AVI), and knowledge informed AVI (i.e., KIAVI, and KIAVI2)) and predictive accuracy in stepwise algorithms. For details of the variable selection methods, please see: Li, J., Siwabessy, J., Huang, Z. and Nichol, S. (2019) . Li, J., Alvarez, B., Siwabessy, J., Tran, M., Huang, Z., Przeslawski, R., Radke, L., Howard, F., Nichol, S. (2017). .

A Toolbox for Conditional Inference Trees and Random Forests

Additions to 'party' and 'partykit' packages : tools for the interpretation of forests (surrogate trees, prototypes, etc.), feature selection (see Gregorutti et al (2017) , Hapfelmeier and Ulm (2013) , Altmann et al (2010) ) and parallelized versions of conditional forest and variable importance functions. Also modules and a shiny app for conditional inference trees.

Variable Importance Measures for Multivariate Random Forests

Calculates two sets of post-hoc variable importance measures for multivariate random forests. The first set of variable importance measures are given by the sum of mean split improvements for splits defined by feature j measured on user-defined examples (i.e., training or testing samples). The second set of importance measures are calculated on a per-outcome variable basis as the sum of mean absolute difference of node values for each split defined by feature j measured on user-defined examples (i.e., training or testing samples). The user can optionally threshold both sets of importance measures to include only splits that are statistically significant as measured using an F-test.