Multilinear Partial Least Squares

Multilinear Partial Least Squares is a multi-way extension of standard PLS. See also: Multiway calibration. Multilinear PLS, Bro 96.

This implementation extends the PLS2 algorithm to threeway input and thus works on multitarget Y data.

Parameters

Parameter Name	Default Value	Description
numComponents	10	Number of components of the loading matrices.
standardizeY	true	Whether to standardize the Y target matrix

Example Code

// Get data
double[][][] xtrain = ... // e.g. load data of shape (I_train x J x K)
double[][][] xtest = ... // e.g. load data of shape (I_test x J x K)
double[][] ytrain = ... // e.g. load data of shape (I_train x M)
double[][] ytest = ... // e.g. load data of shape (I_test x M)
Tensor Xtr = Tensor.create(xtrain);
Tensor Xte = Tensor.create(xtest);
Tensor Ytr = Tensor.create(ytrain);
Tensor Yte = Tensor.create(ytest);

// Setup model
int nComponents = ... // Choose a number of components F for the loading matrices
MultiLinearPLS npls = new MultiLinearPLS();
npls.setNumComponents(nComponents);

// Build and test model
npls.build(Xtr, Ytr);
Tensor Ypred = npls.predict(Yte);
double mse = MathUtils.meanSquaredError(Yte, Ypred);

// Usage as a filter/feature-transformation
Tensor transformedXte = npls.filter(Xte)