Extremal Random Forests – Numerical experiments

The goal of this repository is to reproduce the numerical experiments from [1].

If you use this software in your work, please cite it using the following metadata.

@misc{gnecco2023extremal,
      title={Extremal Random Forests}, 
      author={Nicola Gnecco and Edossa Merga Terefe and Sebastian Engelke},
      year={2023},
      eprint={2201.12865},
      archivePrefix={arXiv},
      primaryClass={stat.ME}
}

Table of Contents

• Directory Structure
• Dependencies
• Clone Repository
• Instructions
• Contributors
• License
• References

Directory Structure

Below is a detailed description of the folder and file structure in the erf-numerical-results repository.

• Data/: Contains the US Census datasets used in the experiments from [2].
• R/: R functions and settings for running numerical experiments (e.g., data analysis, visualization, etc.)
• configs/: JSON configuration files for setting up and running numerical experiments.
• figures/: Generated figures and graphs from the numerical experiments.
• main/: Main scripts for reproducing the numerical experiments from the paper.
• output/: Intermediate results from the numerical experiments.

Dependencies

• R: version (4.0.2)

Installing R Requirements

1. Ensure R is installed: Make sure you have R installed on your system by running in the terminal:

R --version

If you don't have R installed please download R and install it.

2. Install the requirements: Move to the main directory and launch the R script dependencies.R by running in the terminal:

cd R-code
Rscript --vanilla main/dependencies.R

Clone Repository

The repository works with git lfs [https://git-lfs.com/]. After you install git lfs on your system, you can clone the GitHub directory by typing the following in your command line.

git lfs clone https://github.com/nicolagnecco/erf-numerical-results.git

Instructions

All the code to reproduce results in the article is located in the main folder. Below we provide a detailed breakdown of each file.

Note

As an example consider the third row of the table referring to sec_4_1-experiments.R. The file produces Figure 3 of the paper. Its configuration file exp_sec_4_1.json includes the parameter to run the simulation, including the number of parallel R instance (see n_workers). It produces intermediate results, which are saved in the output/ folder, and it produces figures, which are saved in the figures/ folder. For this file, you can plot the results without running the simulation by setting RUN_SIMULATION <- FALSE inside the file. Alternatively, you can run the simulation and plot the results by setting RUN_SIMULATION <- TRUE. The estimated runtime is approximately 15 mins.

Warning

The estimated runtime is based on the number of parallel R instances set in the corresponding config file (see the parameter n_workers there). Running several R instances in parallel (e.g., n_workers > 20) on a local computer might not be feasible.

Main Folder Breakdown

File Name	Description	Config	Intermediate Results?	Figures?	Only Plot Results?	Estimated Runtime
sec_3-generative_model.R	Produce Figure 1		no	yes	no	< 5 mins
sec_3-cv_lambda.R	Produce Figure 2	exp_sec_3_lambda.json exp_sec_3_cv_lambda.json	yes	yes	yes	~ 30 mins
sec_4_1-experiments.R	Produce Figure 3	exp_sec_4_1.json	yes	yes	yes	~ 15 mins
sec_4_1-boxplots.R	Produce Figure 4	exp_sec_4_1_boxplots.json	yes	yes	yes	~ 30 mins
sec_5-wage_analysis.R	Run US Census analysis without plots		yes	no	no	~ 15 mins
sec_5-wage_plots.R	Plots results for US Census analysis, including Figures 5, 6, 7, and Figures 13, 14, 15, 16 (appendix)		no	yes	yes	< 5 mins
sec_5-wage.R	Runs sec_5-wage_analysis.R and then sec_5-wage_plots.R		yes	yes	yes	~ 15 mins
sec_app-similarity_weights.R	Produce Figure 8 (appendix)		no	yes	yes	< 5 mins
sec_3-cv_kappa.R	Produce Figure 9 (appendix)	exp_sec_3_kappa.json exp_sec_3_cv_kappa.json	yes	yes	yes	~ 30 mins
sec_4_2-boxplots.R	Produce Figure 10 (appendix)	exp_sec_4_2_boxplots.json	yes	yes	yes	~ 90 mins
sec_4_2-hill.R	Produce Figure 11 (appendix)	exp_sec_4_2_hill_vs_erf.json	yes	yes	yes	~ 15 mins
sec_4_1_bias_var-experiments.R	Produce Figure 12 (appendix)	exp_sec_4_1_bias_var.json	yes	yes	yes	~ 15 mins

Contributors

• Nicola Gnecco
• Edossa Merga Terefe
• Sebastian Engelke

License

This project is licensed under the MIT License - see the LICENSE file for details.

References

[1]: Nicola Gnecco, Edossa Merga Terefe, and Sebastian Engelke. 2022. "Extremal Random Forests." arXiv Preprint [https://arxiv.org/abs/2201.12865].

[2]: Joshua D. Angrist, Victor Chernozhukov, and Iván Fernández-Val. 2009. "Replication data for: Quantile Regression under Misspecification, with an Application to the U.S. Wage Structure." Harvard Dataverse, V1. DOI: 10.7910/DVN/JNEOLQ.