This repository contains the code used to generate results shown in Russeil et al., 2024, Multi-View Symbolic Regression.

Content

Files:

• mvsr.py: MvSR basic implementation.
• generate_data.py: generates the artificial benchmark dataset
• analysis.py: Run SR/MvSR on artificial benchmark. Allow to refit and evaluate.
• run_all.sh: Run the main analysis for every setup presented in the paper.
• results.py: Read results and aggregate them into a table.
• plots.py: Generate plots from the aggregated table.

Folders:

• real_data : The real datasets from chemistry, finance and astrophysics
• For each data set, we provide a notebook with the specific setups used to generate the parametric functions presented in the paper.

To run the code, in addition to the dependencies listed in the requirement.txt file, it requires two additional setups.

First requirement

The iminuit version we use has been modify to fix an occuring error from version 2.24. The cost.py file was modified so that:

Line 1827 becomes

self._ndim = x.shape[0]

And line 1836 becomes

 x = self._masked.T[: self._ndim]

Second requirement

The pyoperon version used includes some adaptations to make MvSR possible. Therefore the following wheel should be used : https://github.com/erusseil/MvSR-analysis/files/14389428/pyoperon-wheel.zip (compatible with Python 3.8, 3.9, 3.10, 3.11)

After that you can just:

pip install wheel-filename

Installation instructions

This step-by-step will ensure the installation of the correct version of the adapted Operon:

• Install PyEnv (https://github.com/pyenv/pyenv?tab=readme-ov-file#installation)
• Run these commands:

pyenv install 3.11
pyenv shell 3.11
python3.11 -m venv mvsrenv
source mvsrenv/bin/activate
pip3.11 install wheels/pyoperon-0.3.6-cp311-cp311-linux_x86_64.whl
pip3.11 install -r requirement.txt

• Edit the file mvsrenv/lib/python3.11/site-packages/iminuit/cost.py and replace
  • line 1827 with self._ndim = x.shape[0]
  • line 1836 with x = self._masked.T[: self._ndim]