Sparse SVD (svds) in SciPy

Introduction

Sparse SVD helps find important patterns in big, mostly empty data. It works fast and saves memory.

You have a large matrix with many zeros, like user ratings for movies.

You want to reduce data size but keep main information.

You need to find main features or topics in text data stored as sparse matrix.

You want to speed up calculations on big sparse data.

You want to do recommendation systems or clustering on sparse data.

Syntax

SciPy

from scipy.sparse.linalg import svds

u, s, vt = svds(A, k=k)

A is your sparse matrix (usually in CSR or CSC format).

k is the number of singular values and vectors you want.

Examples

Compute 2 largest singular values and vectors of a small sparse matrix.

SciPy

from scipy.sparse import csr_matrix
from scipy.sparse.linalg import svds

A = csr_matrix([[0, 0, 3], [4, 0, 0], [0, 5, 0]])
u, s, vt = svds(A, k=2)

Get the single largest singular value from sparse matrix A.

SciPy

u, s, vt = svds(A, k=1)
print(s)

Sample Program

This code creates a sparse matrix with mostly zeros. Then it finds the 2 biggest singular values and their vectors using svds. It prints these values and vectors.

SciPy

from scipy.sparse import csr_matrix
from scipy.sparse.linalg import svds
import numpy as np

# Create a sparse matrix with many zeros
A = csr_matrix([
    [0, 0, 3, 0],
    [4, 0, 0, 0],
    [0, 5, 0, 0],
    [0, 0, 0, 6]
])

# Compute 2 largest singular values and vectors
u, s, vt = svds(A, k=2)

print("Singular values:", s)
print("Left singular vectors (u):\n", u)
print("Right singular vectors (vt):\n", vt)

Important Notes

Make sure your matrix A is in sparse format like CSR or CSC for best speed.

The svds function returns singular values in ascending order, smallest to largest.

Choosing k too large can slow down the calculation or cause errors.

Summary

Sparse SVD finds main patterns in big, mostly empty data efficiently.

Use svds from scipy.sparse.linalg with sparse matrices.

It returns singular values and vectors that help understand or reduce data.