as_factoextra: Build a factoextra-ready object from pre-computed coordinates

In factoextra: Extract and Visualize the Results of Multivariate Data Analyses

Build a factoextra-ready object from pre-computed coordinates

Description

as_factoextra_pca() wraps pre-computed individual (and, optionally, variable) coordinates into an object that the fviz_pca family (fviz_pca_ind(), fviz_pca_var(), fviz_pca_biplot()), fviz_eig, fviz_contrib and fviz_cos2 can plot directly. Read more: Principal Component Analysis (PCA) in R: Compute, Visualize & Interpret.

It lets you apply factoextra's visualizations to the output of an eigenvalue-based dimension reduction - for example stats::cmdscale(), ape::pcoa(), vegan::rda()/cca(), a tidymodels recipe/workflow with step_pca() (see the methods below), or a custom analysis - without having to write a dedicated backend. You bring the coordinates; factoextra draws the biplot, scree plot, contributions and cos2.

The scree plot, contributions and cos2 assume real eigenvalues, so this constructor is for PCA-family results. Non-linear embeddings (UMAP, t-SNE) have no eigenvalues; plot their coordinates directly (a scree/loadings display would be meaningless for them).

Usage

as_factoextra_pca(ind.coord, ...)

## Default S3 method:
as_factoextra_pca(
  ind.coord,
  var.coord = NULL,
  eig = NULL,
  ind.cos2 = NULL,
  ind.contrib = NULL,
  var.cos2 = NULL,
  var.contrib = NULL,
  var.cor = NULL,
  scale.unit = FALSE,
  ...
)

## S3 method for class 'recipe'
as_factoextra_pca(ind.coord, ...)

## S3 method for class 'workflow'
as_factoextra_pca(ind.coord, ...)

Arguments

ind.coord	the object to convert. For the default method, individual (observation) coordinates: a numeric matrix or data frame with one column per dimension (the "scores"). For the recipe / workflow methods, a prepped recipe or a fitted workflow whose PCA is done with recipes::step_pca() (the scores, loadings and eigenvalues are then extracted for you). Required.
...	passed to methods (unused by the default method).
var.coord	optional variable coordinates / loadings: a numeric matrix or data frame with one column per dimension. Supplying it enables fviz_pca_var() and fviz_pca_biplot().
eig	optional numeric vector of eigenvalues (length \ge number of dimensions). Used by the scree plot and to label the axes with the percentage of explained variance. When NULL (default) it is set to the variance of each coordinate column (the natural definition of a PCA eigenvalue).
ind.cos2, ind.contrib, var.cos2, var.contrib, var.cor	optional pre-computed quality (cos2), contribution and (variable) correlation matrices, with the same dimensions as the corresponding coordinates. When omitted they are derived from the coordinates (see Details).
scale.unit	logical. If TRUE, the variable coordinates are treated as correlations and the correlation circle is drawn by fviz_pca_var(). Default is FALSE.

Details

When cos2/contrib are not supplied they are computed from the coordinates:

• contrib = 100 * coord^2 / colSums(coord^2) - the exact contribution of each element to each dimension.

• cos2 = coord^2 / rowSums(coord^2) - the quality of representation within the supplied dimensions. This equals the true cos2 only when all components are provided; with a truncated set of dimensions it is the quality restricted to that sub-space. Pass ind.cos2/var.cos2 explicitly if you have the exact values.

tidymodels (recipe / workflow). The recipe and workflow methods extract a PCA fitted with recipes::step_pca() through the public recipes/workflows API: the scores from the baked training data (or, for a fitted workflow, workflows::extract_mold()), the loadings from tidy(step, type = "coef"), and the full set of eigenvalues from tidy(step, type = "variance") (so the scree plot and axis percentages are honest even when num_comp keeps only a few components). Variable coordinates are loading times the square root of the eigenvalue. Exact variable-component correlations and cos2 are recovered from the full PCA inertia when every PCA input is provably centered. When those metrics cannot be recovered (for example, for a bare step_pca() with no centering or a zero-inertia variable), their entries in get_pca_var() are NULL. Scores, eigenvalues, variable coordinates and contributions are still returned, so the individual, scree, variable-arrow and biplot displays remain available; correlation/cos2-dependent displays fail with an explicit unavailable-metric error. The correlation circle is omitted and a warning explains how to recover the metrics. scale.unit is set to TRUE only when every PCA input is both centered and unit-scaled at the PCA boundary; partial scaling therefore does not draw a correlation circle. For fully normalized data, variable coordinates, correlations, cos2 and contributions, and the eigenvalue percentages match a full FactoMineR::PCA() regardless of how many components step_pca() keeps; the individual cos2 is computed over the retained components (it equals the full-space cos2 only when all components are kept). The two-dimensional plots (fviz_pca_ind(), fviz_pca_var(), fviz_pca_biplot()) need num_comp >= 2. The recipe must be prepped, and its single PCA step must be an unweighted step_pca() and must be the final recipe step. Case-weighted PCA is rejected because the adapter cannot yet propagate those weights into individual contributions. Later steps could transform the baked/mold PCA scores without updating the fitted PCA loadings or eigenvalues, so such recipes fail explicitly instead of returning internally inconsistent geometry. Non-linear embedding steps such as step_umap() have no eigenvalues and are rejected with a message. For the loadings display of a recipe PCA see also learntidymodels::plot_top_loadings().

Value

An object of class c("factoextra_pca", "list") holding the standardized ind (and var) results and eigenvalues, ready for the fviz_pca_*() functions.

Author(s)

Alboukadel Kassambara [e-mail]

See Also

For UMAP / t-SNE embeddings, which have no eigenvalues, use fviz_umap / fviz_tsne instead. Online tutorial: Principal Component Analysis (PCA) in R: Compute, Visualize & Interpret.

Examples

# 1. Bring your own coordinates: classical MDS (cmdscale) -> factoextra
d <- dist(scale(mtcars))
mds <- cmdscale(d, k = 3)
obj <- as_factoextra_pca(ind.coord = mds)
fviz_pca_ind(obj, repel = TRUE)
fviz_eig(obj)

# 2. Round-trip a prcomp result through the constructor (biplot)
pca <- prcomp(iris[, -5], scale. = TRUE)
obj2 <- as_factoextra_pca(
  ind.coord = pca$x,
  var.coord = sweep(pca$rotation, 2, pca$sdev, "*"),
  eig       = pca$sdev^2,
  scale.unit = TRUE
)
fviz_pca_biplot(obj2, label = "var", col.ind = "steelblue")

# 3. A tidymodels recipe PCA (step_pca) -> factoextra biplot + honest scree
if (requireNamespace("recipes", quietly = TRUE)) {
  library(recipes)
  rec <- recipe(~ ., data = iris[, 1:4]) |>
    step_normalize(all_numeric_predictors()) |>
    step_pca(all_numeric_predictors(), num_comp = 4)
  obj3 <- as_factoextra_pca(prep(rec))
  fviz_pca_biplot(obj3, label = "var")
  fviz_eig(obj3, addlabels = TRUE)
}

factoextra documentation built on July 24, 2026, 9:06 a.m.