nipals-pca/src/pca.js

A NIPALS implementation of PCA that handles missing values

const { Matrix } = require('ml-matrix');

class PCA {
    constructor(dataset, A) {
        this.dataset = dataset;
        this.A = A;
        this.T = new Matrix(this.dataset.matrix.rows, this.A);
        this.P = new Matrix(this.dataset.matrix.columns, this.A);
    }

    fit() {

        let dCrit = 1e-20
        let maxIter = 1000
        let iter = 0
        let diff = 1

        for (let a = 0; a < this.A; a++) {
            iter = 0
            diff = 1

            this.T.setColumn(a, this.dataset.matrix.getColumn(0))

            while (diff > dCrit && iter < maxIter) {
                diff = this.nipals(a)

                iter += 1
            }

            this.dataset.matrix.subtract(this.T.getColumnVector(a).mmul(this.P.getColumnVector(a).transpose()))
            this.dataset.setZeroes()
        }
    }

    nipals(a) {

        let t = this.T.getColumnVector(a)

        let p = this.dataset.matrix.transpose().mmul(t)

        if(this.dataset.missing){
            let pcorr = t.clone().pow(2).transpose().mmul(this.dataset.xmask).transpose();
            let zvals = pcorr.getColumn(0).map((v, i) => v == 0 ? i : -1).filter(idx => idx >= 0);

            p.divColumnVector(pcorr);

            zvals.map(zidx=>p.set(zidx,0,0));
        }

        p.div(p.norm())

        let tnew = this.dataset.matrix.mmul(p)

        if(this.dataset.missing){
            tnew.divColumnVector(this.dataset.xmask.mmul(p.clone().pow(2)))
        }

        let diff = t.subColumnVector(tnew).pow(2).sum();

        this.T.setColumn(a, tnew)
        this.P.setColumn(a, p)

        return diff;
    }
}

module.exports.PCA = PCA;