Code Supplement for 'Fast Estimation of Partial Dependence Functions using Trees'

The following packages were cloned and modified for the simulations

• shap, cloned 14/10/2024, SHA: 215c58525127761208dd2c9754bd468a0295f537 on master
• glex, cloned 19/11/2024, SHA: bb2be47ca2eb2a657095ecfd1b51877dfe0debe9 on master

The folder codeforsimulations is a git repository, version history can be viewed with git log

Reproducibility

The seed used in the Figures for the main article is not the seed used in the simulations here, so therefore there might be slight variations in Figures 1 and 4 due to seed randomness.

To reproduce Figures 1, 2 and 4, follow these steps:

• Change working directory to simulation_glex
• Ensure R (4.4.1) is installed with: scico, devtools, data.table, tidyverse, reshape2, patchwork, future, future.apply, doFuture, xgboost, mvtnorm, mlr3verse
• Run run.r to generate complete_res.RData (takes a long time, suggested to run this on a cluster)
• Run plot_fig.r in the folders figure1, figure2, figure4 to generate the plots. The plot will use the data from complete_res.RData

To reproduce Figure 3, follow these steps:

• Change working directory to simulation_runtime_shap_vs_glex
• Ensure R (4.4.1) is installed with: scico, devtools, xgboost, data.table, xgboost, tidyverse, reshape2, bench
• Run generate_dat_and_model.r to generate dataset of 10000 observations and XGBoost model with 20 trees
• Run benchmark_fastpd.r to get the FastPD benchmarks (takes a long time, suggested to run this on a cluster)
• Ensure Python >=3.12 is installed
• Make a Python virutal environment using python3.12 -m venv venv
• Activate the venv source venv/bin/activate
• Install the modified SHAP package pip install ../shap
• Install XGBoost if not exists pip install xgboost
• Run python benchmark_shap.py to get the SHAP benchmarks (takes a long time, suggested to run this on a cluster)
• Change working directory to figure3
• Run plot_runtime.r to get the benchmark plots

SHAP package modifications

• Removed `docs, data, javascript, notebooks, scripts
• Modified shap/explainers/_tree.py

glex package modifications

• Added src/recurse_fastpd.cpp
• Added R/fastpd.R
• Added auxillary functions in glex.R to call the FastPD implementation