PLS and CCA

Algorithm overview

Partial least squares finds latent directions that predict a target block, while canonical correlation analysis finds directions with high cross-block correlation. Both help with correlated, multivariate features.

JavaScript implementation

PLSRegression uses mode-A NIPALS with regression deflation. CCA uses mode-B weights with canonical deflation.

Interactive PLS regression playground

Change the latent-component count and data distribution below. The visible first target and holdout metrics come from a live multi-output CrossDecomposition.PLSRegression fit.

Quick start

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

const X = [[0, 1], [1, 2], [2, 3], [3, 5]];
const Y = [[1], [2], [4], [7]];
const model = new CrossDecomposition.PLSRegression({ nComponents: 2 });
model.fit(X, Y);
console.log(model.predict([[4, 6]]));

Detailed API reference

Both expose weights, loadings, rotations, iteration counts, transform, and predict.