lihaicheng7003 / mrmr

Python implementation of mRMR (minimum-Redundancy-Maximum-Relevance) for feature selection.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

mrmr

mrmr (Minimum-Redundancy-Maximum-Relevance) is a "minimal optimal" feature selection algorithm, meaning that it seeks to find a feature set giving the best possible predictive performance, given a fixed number of features.

How to install

You can install mrmr in your environment via:

pip install git+https://github.com/smazzanti/mrmr

How to use

You have a dataframe composed of numeric variables (X) and a series which is your (binary or multiclass) target variable (y). You want to select K features such that they are maximally relevant, but also as little redundant as possible with each other.

from mrmr import mrmr_classif
from sklearn.datasets import make_classification

# create some data
X, y = make_classification(n_samples = 1000, n_features = 50, n_informative = 10, n_redundant = 40)
X = pd.DataFrame(X)
y = pd.Series(y)

# use mrmr classification
selected_features = mrmr_classif(X, y, K = 10)

Note: the output of mrmr_classif is a list containing K selected features. This is a ranking, therefore, if you want to make a further selection, take the first elements of this list.

Reference

For an easy-going introduction to MRMR, read my article on Towards Data Science: “MRMR” Explained Exactly How You Wished Someone Explained to You.

Also, this article describes an example of MRMR used on the world famous MNIST dataset: Feature Selection: How To Throw Away 95% of Your Data and Get 95% Accuracy

MRMR was born in 2003, this is the original paper: Minimum Redundancy Feature Selection From Microarray Gene Expression Data.

Since then, it has been used in many practical applications, due to its simplicity and effectiveness. For instance, in 2019, Uber engineers published a paper describing how they implemented MRMR in their marketing machine learning platform Maximum Relevance and Minimum Redundancy Feature Selection Methods for a Marketing Machine Learning Platform.

About

Python implementation of mRMR (minimum-Redundancy-Maximum-Relevance) for feature selection.

License:GNU General Public License v3.0


Languages

Language:Jupyter Notebook 88.7%Language:Python 11.3%