Generalized Random Forests Using Fixed-Point Trees

Authors: David Fleischer, David A. Stephens, Archer Y. Yang

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate that our method achieves a speedup of multiple times over standard GRFs without compromising statistical accuracy. Experiments on both simulated and real-world data validate our approach. Our findings suggest that the proposed method is a scalable alternative for localized effect estimation in machine learning and causal inference applications.
Researcher Affiliation Academia 1Department of Mathematics and Statistics, Mc Gill University, Montreal, Canada 2Mila Quebec AI Institute, Montreal, Quebec, Canada. Correspondence to: Archer Y. Yang <EMAIL>.
Pseudocode Yes Algorithm 1 The fixed-point tree algorithm... Algorithm 2 Stage I GRF-FPT: Training a generalized random forest using fixed-point trees... Algorithm 3 GRF-FPT: Estimates of θ (x)
Open Source Code Yes We implement the GRF-FPT algorithm in a fork of grf (Tibshirani et al., 2024) available at https://github.com/dfleis/grf... Code and data for reproducing all experiments and figures are available at https://github.com/dfleis/grf-experiments.
Open Datasets Yes The data, first appearing in Kelley Pace & Barry (1997), contains 20,640 observations of housing prices taken from the 1990 California census... can be directly obtained from the Carnegie Mellon Stat Lib repository (https://lib.stat.cmu.edu/datasets/houses.zip).
Dataset Splits Yes Throughout our experiments we use subsampling ratio s/n = 0.5...To assess estimation accuracy, we evaluate the mean squared error (MSE) of ŷ(x) across 50 replications of the model and testing on a separate set of 5,000 observations...We varied the subsampling ratio s/n {0.25, 0.50, 0.75} under VCM Setting 3 over a forest of 10 trees carried out using GRF-FPT2 and GRF-grad.
Hardware Specification No No specific hardware details (like GPU/CPU models, memory) are mentioned in the paper, only that the algorithm is implemented in an R package.
Software Dependencies Yes We implement the GRF-FPT algorithm in a fork of grf (Tibshirani et al., 2024) available at https://github.com/dfleis/grf. R package version 2.4.0.
Experiment Setup Yes Throughout our experiments we use subsampling ratio s/n = 0.5...All versions fit a forest of 2000 trees, the default settings of the original R implementation (Tibshirani et al., 2024), a subsample ratio of 0.5, and a target minimum node size of 5 observations.