Implementation:Scikit learn Scikit learn BenchPlotNMF

Knowledge Sources	Scikit_learn Scikit-learn Docs
Domains	Machine Learning, Benchmarking
Last Updated	2026-02-08 15:00 GMT

Overview

Concrete tool for benchmarking Non-Negative Matrix Factorization (NMF) solvers provided by scikit-learn.

Description

This benchmark script compares multiple NMF solver implementations including coordinate descent, multiplicative update, and a projected gradient variant (_PGNMF). It measures reconstruction error (beta divergence) and runtime across various NMF configurations. The script includes a custom projected gradient NMF implementation for historical comparison purposes and uses TF-IDF vectorized text data for evaluation.

Usage

Use this benchmark to compare the convergence speed and accuracy of different NMF solver strategies available in scikit-learn, especially when choosing between coordinate descent and multiplicative update solvers.

Code Reference

Source Location

• Repository: scikit-learn
• File: benchmarks/bench_plot_nmf.py

Signature

# Key internal functions
def _nls_subproblem(X, W, H, tol, max_iter, alpha=0.0, l1_ratio=0.0, sigma=0.01, beta=0.1)
def _norm(x)

# Main NMF class used
from sklearn.decomposition import NMF

Import

from sklearn.decomposition import NMF
from sklearn.decomposition._nmf import _beta_divergence, _check_init, _initialize_nmf

I/O Contract

Inputs

Name	Type	Required	Description
X	array-like	Yes	Input non-negative data matrix (typically TF-IDF features)
W	array-like	Yes	Basis matrix (n_samples, n_components)
H	array-like	Yes	Coefficient matrix (n_components, n_features)
tol	float	Yes	Tolerance for stopping condition
max_iter	int	Yes	Maximum number of iterations

Outputs

Name	Type	Description
Console output	text	Beta divergence and timing for each solver
Plot	matplotlib figure	Convergence plots comparing NMF solvers

Usage Examples

Basic Usage

from sklearn.decomposition import NMF
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.datasets import fetch_20newsgroups

newsgroups = fetch_20newsgroups(subset='train')
vectorizer = TfidfVectorizer(max_features=1000)
X = vectorizer.fit_transform(newsgroups.data)

nmf = NMF(n_components=10, solver='cd', max_iter=200, random_state=42)
W = nmf.fit_transform(X)
H = nmf.components_

Related Pages

• Principle:Scikit_learn_Scikit_learn_Dimensionality_Reduction