NMF

Algorithm overview

Non-negative Matrix Factorization approximates X ≈ W × H while keeping every factor non-negative. The additive representation is useful for topic features, parts-based image representations, and interpretable latent components.

JavaScript implementation

@kanaries/ml uses multiplicative Frobenius-loss updates and supports random and NNDSVD initializations. It runs in browser or Node.js and exposes both learned components and reconstruction error.

Quick start example

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

const nmf = new Decomposition.NMF({ nComponents: 2, init: 'nndsvda', randomState: 0 });
const W = nmf.fitTransform([[1, 2, 0], [0, 1, 3], [1, 3, 3]]);
const approximation = nmf.inverseTransform(W);
console.log({ W, approximation });

Detailed API reference

Options: nComponents, init: 'random' | 'nndsvd' | 'nndsvda' | 'nndsvdar', maxIter, tol, alphaW, alphaH, l1Ratio, and randomState.

Methods: fit, transform, fitTransform, inverseTransform. Learned fields: components, reconstructionErr, and nIter. Input must be a finite, rectangular, non-negative matrix.