PCA

Algorithm overview

Linear dimensionality reduction using Singular Value Decomposition of the data to project it to a lower dimensional space. The input data is centered but not scaled for each feature before applying the SVD.

PCA compresses correlated numeric features into orthogonal components while preserving maximal variance.

This algorithm is especially useful when:

• Feature dimensions are high and training speed needs improvement.
• You need 2D/3D projections for exploratory visualization.
• Multicollinearity hurts stability in downstream supervised models.

JavaScript implementation

@kanaries/ml gives frontend and full-stack teams a JavaScript PCA implementation for feature compression, denoising, and 2D or 3D visualization. That is useful when your embeddings or tabular features are already flowing through a JS application and you want dimensionality reduction without switching runtimes.

In practice, PCA is often one of the most useful browser-friendly ML utilities because it helps simplify data before charting, clustering, or downstream modeling.

Quick start

from sklearn.decomposition import PCA

X = [[1, 2, 3], [2, 3, 4], [3, 5, 7], [4, 6, 8]]

pca = PCA(n_components=2)
embedding = pca.fit_transform(X)

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

const X = [[1, 2, 3], [2, 3, 4], [3, 5, 7], [4, 6, 8]];

const pca = new Decomposition.PCA(2);
const embedding = pca.fitTransform(X);

Quick JavaScript example

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

const X = [[1, 2, 3], [2, 3, 4], [3, 5, 7], [4, 6, 8]];

const pca = new Decomposition.PCA(2);
const embedding = pca.fitTransform(X);

Detailed API reference

Algorithm

The covariance matrix of the centered samples is decomposed by power iteration. The resulting eigenvectors form the principal components sorted by explained variance.

constructor(nComponents: number | null = null)

Parameters

• nComponents (number | null, default null): number of principal components to retain. If null, all components are used.

Methods

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

Implementation workflow

1. Standardize continuous features before decomposition.
2. Fit PCA and select component count by explained variance.
3. Use transformed features for visualization or downstream training.

JavaScript deployment notes

• Standardize continuous features before PCA so dominant dimensions do not reflect raw scale alone.
• Use PCA as a preprocessing step before visualization, clustering, or other downstream models.
• In browser tools, PCA is a practical way to reduce feature count before rendering or interactive analysis.