This repository contains all of the code to reproduce the experiments from 'MAPTree: Beating "Optimal" Decision Trees with Bayesian Decision Trees' by Colin Sullivan*, Mo Tiwari*, and Sebastian Thrun.
Our main algorithm is written in C++ and called from Python via Python bindings.
Below, we have instructions on how to reproduce all of our results. If you have a question about our code, please submit a Github issue.
Our code only supports Python 3.10.
Note that our dependencies require installing the old sklearn package
(not scikit-learn); you may need run the following command to allow this:
export SKLEARN_ALLOW_DEPRECATED_SKLEARN_PACKAGE_INSTALL=True
If you're on an M1 Mac, you'll need to install gosdt from source; see the instructions
here. Afterwards, install
the other dependencies and the maptree Python package with
python -m pip install -r requirements_m1mac.txt
python -m pip install .
If you're not on an M1 Mac, you can install all dependencies and then build the maptree Python package directly:
python -m pip install -r requirements.txt
python -m pip install .
Download all the necessary data with
python setup_data.py
We support running all experiments on a single machine (e.g., a personal laptop) with:
python run_experiment.py
Note that this may take a long time (1-2 weeks), depending on your hardware.
We have also included a script to run the experiments on a cluster via SLURM, which can be invoked with:
sbatch script.slurm
Note that this may take some time (1-2 days) to complete, depending on your hardware, and that the
parameters of the script.slurm file may need to be modified according to your cluster setup.
You can plot the results of all experiments with:
python plot_results.py
The plots can be found in experiments/results/figures.
Hyperparameters for all the experiments can be found in the following files:
experiments/globals.py
experiments/experiments/fig1/constants.py
experiments/experiments/fig2/constants.py
experiments/experiments/fig3/constants.py