Random projection

Algorithm overview

Random projection reduces dimensionality with a data-independent matrix while approximately preserving pairwise distances. It is useful when deterministic decomposition is too expensive.

JavaScript implementation

@kanaries/ml generates seeded Gaussian or sparse projection matrices in browser and Node.js and exposes the fitted components for serialization. Random projection is a lightweight choice when PCA-style fitting is too expensive or data arrives only at inference time; the sparse variant reduces multiply cost for high-dimensional frontend features.

Quick start

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

const X = [[1, 0, 2], [0, 1, 3], [2, 1, 0]];
const model = new RandomProjection.SparseRandomProjection({ nComponents: 128, randomState: 42 });
const reduced = model.fitTransform(X);
console.log(reduced[0].length);

Detailed API reference

Set nComponents: 'auto' with eps to use the sklearn Johnson-Lindenstrauss bound. Both estimators expose components; computeInverseComponents enables approximate inverseTransform.