Types

NCPLS relies on two main container types in the StatsAPI.fit and StatsAPI.predict workflow. NCPLSModel stores the fitting specification: preprocessing options, the number of extracted components, and the switches that control the multilinear loading-weight branch. NCPLSFit is the full fitted model returned by fit; it contains regression coefficients, predictor scores, multilinear weight information, preprocessing statistics, and the labels or axis metadata retained from model fitting.

A third container, NCPLSFitLight, stores only the subset of fitted-model information needed for fast internal prediction during cross-validation. It is mainly an internal optimization and is usually not accessed directly.

The package provides getters for quantities that are commonly useful in downstream work. For fitted models, the most important ones are coef, fitted, residuals, xscores, predictoraxes, responselabels, samplelabels, sampleclasses, xmean, xstd, and ymean. NCPLSModel does not currently have a separate getter layer; its options are accessed directly as fields, for example model.multilinear, model.orthogonalize_mode_weights, or model.ncomponents.

Two stored label vectors are easy to confuse:

• sampleclasses lives on the sample axis and stores one label per training sample,
• responselabels lives on the response axis and stores one label per response column.

These vectors interact only when NCPLS tries to infer a class-response block from a custom response matrix. If the unique sample classes can be matched consistently to a one-hot sub-block of the responses, classification helpers use that sub-block. If not, sampleclasses(mf) remains sample-level metadata for grouping and visualization.

As with the CPPLS documentation, the purpose of this page is to describe the stored containers and the most useful accessors. The modeling choices controlled through fit are discussed on the Fit page, and the downstream interpretation of predictions is discussed on Projection and Prediction.

NCPLSModel

NCPLSModel is the user-facing model specification. It stores:

• the number of latent components to extract,
• predictor preprocessing choices (center_X, scale_X),
• primary-response centering (center_Yprim),
• whether predictor-side weights stay unfolded or are compressed to rank-1 multilinear tensors (multilinear),
• whether multilinear mode vectors are orthogonalized across components (orthogonalize_mode_weights),
• and the PARAFAC control parameters used in the multilinear branch.

The default NCPLSModel() uses multilinear=true, so genuinely multiway predictors are handled through the multilinear weight representation unless you request the unfolded comparison branch explicitly.

NCPLSModel

PredictorAxis

PredictorAxis stores metadata for one non-sample predictor mode. It is useful when your predictor tensor has meaningful coordinates such as time, wavelength, or m/z values, and you want those coordinates retained on the fitted model for downstream plots or inspection.

Provide one PredictorAxis per non-sample predictor mode through the predictoraxes keyword in fit. The stored tuple can then be retrieved with predictoraxes.

PredictorAxis(name, values; unit=nothing)
PredictorAxis(; name, values, unit=nothing)

AbstractNCPLSFit

AbstractNCPLSFit is the common supertype of full and reduced fitted NCPLS models. The generic fitted-model accessors are defined at this level.

coef returns the regression coefficients for the final or requested number of components. These coefficients act on predictors after the fitted model's centering and optional scaling have been applied. xmean, xstd, and ymean expose the stored preprocessing statistics, and ncomponents returns the number of latent components retained in the fitted object.

AbstractNCPLSFit

coef(mf::AbstractNCPLSFit)
coef(mf::AbstractNCPLSFit, ncomps::Integer)

xmean(mf::AbstractNCPLSFit)

xstd(mf::AbstractNCPLSFit)

ymean(mf::AbstractNCPLSFit)

ncomponents(mf::AbstractNCPLSFit)

NCPLSFit

NCPLSFit is the full fitted model returned by fit. In addition to the coefficient tensor, it stores predictor scores, loadings, projection objects, response residuals, and the multilinear diagnostics associated with each extracted component.

The most commonly used fitted-model getters are:

• fitted for fitted responses on the original response scale,
• residuals for response residuals,
• xscores for latent predictor scores,
• responselabels, samplelabels, and sampleclasses for stored labels,
• predictoraxes for retained predictor-axis metadata.

For deeper inspection, the full stored fields remain accessible via dot notation. This is particularly useful for multilinear fits, where mf.W_modes, mf.W_multilinear_relerr, mf.W_multilinear_method, mf.W_multilinear_lambda, mf.W_multilinear_niter, and mf.W_multilinear_converged expose the component-wise PARAFAC approximation details.

NCPLSFit

fitted(mf::NCPLSFit)
fitted(mf::NCPLSFit, ncomps::Integer)

predictoraxes(mf::NCPLSFit)

residuals(mf::NCPLSFit)
residuals(mf::NCPLSFit, ncomps::Integer)

responselabels(mf::NCPLSFit)

sampleclasses(mf::NCPLSFit)

samplelabels(mf::NCPLSFit)

xscores(mf::NCPLSFit)
xscores(mf::NCPLSFit, comp::Integer)
xscores(mf::NCPLSFit, comps::AbstractUnitRange{<:Integer})
xscores(mf::NCPLSFit, comps::AbstractVector{<:Integer})

NCPLSFitLight

NCPLSFitLight is a reduced fitted-model container used mainly in cross- validation, permutation testing, and nested resampling. It stores only what prediction needs: the coefficient tensor and the preprocessing statistics required to restore the predictor and response scales.

Most users will interact with it only indirectly through the resampling helpers.

NCPLSFitLight{TB, TXStat, TYStat}