mrmrs 0.1.3

Ultra fast mRMR algorithm for python

MRMRS

Blazing-fast feature selection using the mRMR (minimum Redundancy - Maximum Relevance) algorithm.

• 🚀 100-500x faster than mrmr-selection
• 🐻‍❄️ Works directly on Polars DataFrames — zero-copy, no conversions needed
• ⚡ Sub-5ms import time — no heavy dependencies, just a Rust binary
• 🦀 Written in Rust with SIMD acceleration and incremental redundancy updates

Installation

uv add mrmrs

or

pip install mrmrs

Quick Start

import polars as pl
from mrmrs import mrmr, Feature

X: pl.DataFrame = pl.read_csv("dataset.csv")
y: pl.Series = pl.read_csv("target.csv")["target"]

selected_features: list[Feature] = mrmr(
    X=X,
    y=y,
    number_of_features=50,
    task_type="classification" # or "regression"
)

selected_feature_names: list[str] = [feature.name for feature in selected_features]
subset_df = df.select(selected_feature_names)

Supported Methods

Relevance

• F-statistic from one-way ANOVA (classification)
• Pearson's correlation coefficient (regression)

Redundancy

• Pearson's correlation coefficient

Development

Requirements

• uv, cargo, maturin

1. create uv venv

uv sync

2. activate uv venv

source .venv/bin/activate

3. install this pkg to this uv venv in release mode

maturin develop --release --uv

alternatively if you enter this repository while you have another venv activated and run the same command maturin will install this pkg to that environment

Running tests

• currently there is no separation between the rust crate and the python bindings so this requries the venv to be activated

cargo test

Benchmarks

Benchmarked against the mrmr-selection package on synthetic data generated with sklearn.datasets.make_classification (1000 samples, classification task). Each configuration was run 5 times, reporting mean +/- std.

mrmrs achieves its speed through Rust with SIMD (scalar fallback when unavailable), zero-copy Polars Arrow memory, and an incremental redundancy update that avoids recomputation each round.

Scaling total features (selecting 50 features)*

Features	mrmrs (s)	mrmr_selection (s)	Speedup
500	0.009 ± 0.001	2.286 ± 0.602	198x
1,000	0.016 ± 0.002	3.181 ± 0.041	234x
5,000	0.069 ± 0.005	18.242 ± 0.042	269x
10,000	0.170 ± 0.005	30.680 ± 0.130	184x
20,000	0.357 ± 0.010	56.938 ± 1.506	155x
50,000	0.963 ± 0.007	136.115 ± 1.191	143x

Scaling selected features (from 5,000 total features)*

Selected (K)	mrmrs (s)	mrmr_selection (s)	Speedup
10	0.048 ± 0.001	4.201 ± 0.112	89x
50	0.070 ± 0.001	18.473 ± 0.218	263x
100	0.097 ± 0.002	36.183 ± 0.076	380x
200	0.149 ± 0.005	71.431 ± 0.622	519x

*On AMD Ryzen 5 5600 6-Core (12 threads), 32 GB RAM

The speedup increases with larger K because mrmrs calculates redundancies incrementally in parallel while mrmr_selection recomputes them from scratch each round.

To reproduce: uv run --package mrmrs-bench python bench/benchmark.py

References

• mrmr-selection
• Numerically stable parallel computation of (co-)variance
• Minimum Redundancy Feature Selection from Microarray Gene Expression Data (o.g. mRMR paper)
• Maximum Relevance and Minimum Redundancy Feature Selection Methods for a Marketing Machine Learning Platform

AI Disclaimer

All library code is handwritten (cause this is a sideproject for fun and I like doing it), claude is useful for writing tests and github actions though