sklearn-discretefirstorder 0.0.2

A scikit-learn compatible discrete first-order method for subset selection.

A Discrete First Order Method for Subset Selection

sklearn-discretefirstorder is a light-weight package that implements a simple discrete first-order method for best feature subset selection in linear regression.

The discrete first-order method is based on the technique described by Berstimas et al. [1]

The package is built on top of the scikit-learn framework and is compatible with scikit-learn methods such as cross-validation and pipelines. I followed the guidelines for developing scikit-learn estimators as outlined in the scikit-learn documentation.

About the project

I created this project first and foremost to learn more about how to build and maintain a Python project. My goal was never to build a state-of-the-art machine learning package.

I picked this topic because I had experimented with different feature selection approaches (including the discrete first-order method implemented here) as part of a grad school class project. However, I never developed a robust, well-documented codebase. I decided to re-implement the simplest technique from my grad school project so that I could focus on key aspects of project development such as proper API design, documentation and testing.

I felt like the scikit-learn framework was appropriate for this ML use-case and, more generally, a good set of guiding principles for my first proper Python package thanks to its clear standards and good documentation.

At the moment, the project is in a very early stage of development, but basic usage is already possible. If time permits, I plan to add more features and improve the documentation.

Installation

To install the package from source, clone this repo and run pip install:

git clone https://github.com/miguelfmc/sklearn-discretefirstorder
cd sklearn-discretefirstorder
pip install .

The package is also available for installation from PyPI!:

pip install sklearn-discretefirstorder

Quick Start

Once you have installed the package you can start using it as follows:

The key estimator in this packages is the discretefirstorder.DFORegressor. You can import it as:

from discretefirstorder import DFORegressor

Easily fit the estimator as follows:

import numpy as np
from discretefirstorder import DFORegressor
X = np.arange(100).reshape(100, 1)
y = np.random.normal(size=(100, ))
estimator = DFORegressor()
estimator.fit(X, y)

For more examples, see the documentation.

Known Issues

This package is still at a very early stage of development. The following issues are known:

• Optimization routines are implemented in Python, which makes them slow.

• At the moment, the package only supports squared error loss minimization but there are plans to include support for absolute error loss minimization.

• At the moment, there is no support for classification problems i.e. logistic regression.

• I’m working on making the package available on conda-forge. Stay tuned for updates!

Contributing

While the project is still in its early stages, contributions are welcome!

To contribute, please fork the repo and clone it to your local machine. Then, create a new branch and make your changes.

I suggest you set-up your local environment with conda and pip:

conda create -n sklearn-discretefirstorder python=3.8
conda activate sklearn-discretefirstorder
pip install -r requirements.txt -r requirements-dev.txt -r requirements-docs.txt -r requirements-test.txt

You can also use conda to install all the dependencies from the environment.yml file:

conda env create -f environment.yml
conda activate sklearn-discretefirstorder

Then, install the package in editable mode:

pip install -e .

License

Distributed under BSD 3-Clause License. See LICENSE for more information.

References

[1]

Dimitris Bertsimas. Angela King. Rahul Mazumder. “Best subset selection via a modern optimization lens.” Ann. Statist. 44 (2) 813 - 852, April 2016. https://doi.org/10.1214/15-AOS1388