recipes step_pca() in R: PCA Feature Reduction for Modeling

The recipes step_pca() function in R adds principal component analysis to a preprocessing pipeline, replacing a set of correlated numeric predictors with a smaller block of uncorrelated components before a model is trained.

step_pca(rec, all_numeric_predictors())                            # default 5 components
step_pca(rec, all_numeric_predictors(), num_comp = 3)              # keep 3 components
step_pca(rec, all_numeric_predictors(), threshold = 0.9)           # keep 90% of variance
step_pca(rec, x, y, z)                                             # named columns only
step_pca(rec, all_numeric_predictors(), prefix = "pc_")            # rename PC1 to pc_1
step_pca(rec, all_numeric_predictors(), keep_original_cols = TRUE) # keep input columns
prep(rec); bake(rec, new_data = NULL)                              # train, then apply
tidy(prep(rec), number = 1, type = "variance")                     # variance per component

What step_pca() does

step_pca() converts a group of numeric predictors into principal components. Principal component analysis finds new axes, ordered by how much variance they capture, that are linear combinations of the original columns. When predictors are correlated, the first few components hold most of the signal, so a model can train on three or four components instead of twenty columns.

Like every recipe step, step_pca() works in two stages. During prep() it runs prcomp() on the selected training columns and stores the rotation matrix, the loadings that map raw columns to components. bake() then projects any data, training or new, onto those stored axes. The output columns are named PC1, PC2, and so on, and the original predictors are dropped.

step_pca() syntax and arguments

step_pca() is a transformation step sized by either a component count or a variance target. You select columns directly or, far more often, with the tidyselect helper all_numeric_predictors(), since PCA is only defined for numeric data.

Two arguments control the output width. num_comp keeps a fixed number of components, while threshold keeps however many are needed to capture a fraction of total variance; setting threshold overrides num_comp. The options list is forwarded to prcomp(), which centers but does not scale by default. Full argument detail lives in the recipes step_pca reference.

step_pca() examples

Every example follows the prep-then-bake rhythm. The first builds a recipe on mtcars, normalizes the ten numeric predictors, and compresses them into three components. Normalizing first is important and is covered in the pitfalls below.

The ten predictors collapse into PC1, PC2, and PC3, while the outcome mpg passes through untouched. Calling bake(pca_rec, new_data = test_df) on any frame with the same input columns reuses the stored loadings, so train and test always land in the same component space.

With num_comp = 5 the baked frame has six columns: five components plus the outcome. Picking a count is fine when you already know how wide you want the model input. When you would rather target a variance budget, use threshold instead.

The four iris measurements are strongly correlated, so two components already explain more than 95% of their variance and step_pca() stops there. To see the variance each component captures, call tidy() on the prepared recipe with type = "variance".

The number = 2 argument points at the second step in the recipe, step_pca(). The table confirms PC1 and PC2 together carry about 96% of the variance, which is why the threshold = 0.95 recipe retained exactly those two. Use type = "coef", the default, to inspect the loadings instead.

step_pca() vs other reduction steps

step_pca() is one of several recipes steps that shrink a wide predictor set, and they leave different data behind. Choosing wrong either keeps redundancy or discards structure the model needs.

Reach for step_pca() when many predictors move together and you accept anonymous components in exchange for a compact, decorrelated model input. Choose step_corr() when interpretability matters and you want to keep real column names. Use step_kpca() or step_ica() only when the linear, variance-ranked view of PCA does not fit the problem.

Common pitfalls

Three mistakes account for most step_pca() confusion.

1. Skipping normalization. step_pca() centers but does not scale by default, so a predictor measured in thousands dominates the first component over one measured in fractions. Add step_normalize() before step_pca(), or pass options = list(scale. = TRUE).
2. Selecting non-numeric columns. PCA is numeric-only. Use all_numeric_predictors() rather than all_predictors(), otherwise factor or character columns trigger an error during prep().
3. Expecting interpretable outputs. Components are blends of every input column. If your report needs named predictors, step_pca() is the wrong tool; use step_corr() or step_select() instead.

Try it yourself

Related recipes functions

step_pca() rarely appears alone in a recipe. These steps commonly sit alongside it:

• step_normalize() centers and scales predictors, the recommended step before step_pca().
• step_corr() drops correlated columns while keeping the survivors interpretable.
• step_kpca() runs kernel PCA for nonlinear predictor structure.
• step_pls() reduces predictors using the outcome, unlike unsupervised PCA.
• step_impute_mean() fills missing values, since PCA cannot handle NA inputs.

FAQ

What does step_pca() do in R?

step_pca() is a recipes step that performs principal component analysis as part of a preprocessing pipeline. During prep() it runs prcomp() on the selected numeric predictors and stores the rotation matrix. During bake() it projects the data onto those axes, replacing the original columns with components named PC1, PC2, and so on. Because the rotation is learned only from training data, the same projection applies cleanly to new data without leakage.

How many components should step_pca() keep?

Use num_comp to keep a fixed count or threshold to keep enough components for a variance fraction, such as threshold = 0.9 for 90%. If you are unsure, set num_comp = tune() and let tune_grid() search for the value that gives the best cross-validated model performance. There is no universal number; it depends on how correlated the predictors are and which model consumes them.

Should I normalize before step_pca()?

Yes. step_pca() centers predictors but does not scale them by default, so a high-variance column on a large numeric scale can dominate the first component. Place step_normalize() before step_pca() in the recipe, or pass options = list(scale. = TRUE). Without one of these, PCA reflects measurement units rather than genuine shared structure.

What is the difference between step_pca() and step_corr()?

step_pca() replaces a correlated group of predictors with new orthogonal components, so the output columns are anonymous blends of the originals. step_corr() instead deletes redundant columns and keeps the survivors exactly as they were, with their original names. Use step_pca() for maximum compression and a decorrelated model input, and step_corr() when you need to keep interpretable, named predictors.