Truncated SVD

Algorithm overview

Dimensionality reduction using truncated Singular Value Decomposition of the data. Unlike PCA, the input data is not centered before decomposition.

TruncatedSVD reduces dimensionality of large sparse matrices without centering, making it practical for text/vectorized data.

This algorithm is especially useful when:

• Your feature matrix is sparse (for example TF-IDF or one-hot encodings).
• You need latent semantic compression before classification or retrieval.
• PCA centering is too expensive or undesirable for sparse data.

JavaScript implementation

@kanaries/ml brings Truncated SVD to JavaScript for workflows that work with large sparse matrices, vectorized text, or one-hot encoded features. This is especially useful in Node.js search, recommendation, and document pipelines where you want latent factors without centering the matrix first.

For product teams working with sparse representations in TypeScript, this keeps feature compression and downstream modeling inside the same JS stack.

Quick start

from sklearn.decomposition import TruncatedSVD

X = [[1, 0, 1], [0, 1, 1], [1, 1, 0], [0, 0, 1]]

svd = TruncatedSVD(n_components=2, random_state=0)
embedding = svd.fit_transform(X)

import { Decomposition } from '@kanaries/ml';

const X = [[1, 0, 1], [0, 1, 1], [1, 1, 0], [0, 0, 1]];

const svd = new Decomposition.TruncatedSVD(2);
const embedding = svd.fitTransform(X);

Quick JavaScript example

import { Decomposition } from '@kanaries/ml';

const X = [[1, 0, 1], [0, 1, 1], [1, 1, 0], [0, 0, 1]];

const svd = new Decomposition.TruncatedSVD(2);
const embedding = svd.fitTransform(X);

Detailed API reference

Algorithm

The algorithm performs power iteration on the uncentered covariance matrix and keeps the top components corresponding to the largest singular values.

constructor(nComponents: number = 2)

Parameters

• nComponents (number, default 2): number of singular vectors to retain.

Methods

• fit(X: number[][]): void
• transform(X: number[][]): number[][]
• fitTransform(X: number[][]): number[][]
• inverseTransform(X: number[][]): number[][]
• getComponents(): number[][]
• getSingularValues(): number[]
• getExplainedVariance(): number[]
• getExplainedVarianceRatio(): number[]

Implementation workflow

1. Build sparse-compatible numeric input matrix.
2. Fit TruncatedSVD with candidate component counts.
3. Evaluate retrieval/classification quality on reduced features.

JavaScript deployment notes

• Prefer Truncated SVD over PCA when working with sparse or high-cardinality vectorized inputs.
• Use it before retrieval, classification, or visualization when you need lower-dimensional latent features.
• It is particularly practical in Node.js pipelines that already generate sparse text or categorical vectors.