recipes step_dummy() in R: Create Dummy Variables
The recipes step_dummy() function in R turns factor and character predictors into numeric 0/1 dummy variables so models can use them. You add it to a recipe(), record the factor levels with prep(), and apply the encoding with bake().
step_dummy(rec, all_nominal_predictors()) # dummy-encode every factor step_dummy(rec, Species) # encode one named column step_dummy(rec, Species, one_hot = TRUE) # keep all levels, no reference step_dummy(rec, all_nominal(), -all_outcomes()) # encode factors, skip the outcome step_dummy(rec, where(is.character)) # encode character columns too prep(rec) |> bake(new_data = NULL) # learn levels, then apply tidy(prep(rec), number = 1) # inspect the retained levels
Need explanation? Read on for examples and pitfalls.
What step_dummy() does in R
step_dummy() replaces a categorical column with a set of binary indicator columns. Most R modeling engines need numeric input, so a factor like Species cannot be fed in directly. step_dummy() converts each factor into one or more 0/1 columns, where a 1 marks membership in a level and a 0 marks absence.
By default it uses treatment contrasts, the same scheme lm() applies behind the scenes. A factor with K levels becomes K-1 dummy columns, and the first level is held back as the reference. A car with 4, 6, or 8 cylinders therefore needs only two columns, because "not 6 and not 8" already identifies the 4-cylinder group.
step_dummy() records every level it sees in the training data and stores them inside the prepped recipe. When you bake() new rows, it reuses that stored level set, so the test data gets exactly the same columns in the same order as training, even if a level happens to be absent from the new batch.step_dummy() syntax and arguments
step_dummy() attaches an encoding operation to a recipe. You pass the recipe first, then the columns to encode, chosen with tidyselect helpers like all_nominal_predictors().
The arguments you will actually touch:
| Argument | Purpose |
|---|---|
recipe |
The recipe object the step is added to. |
... |
Columns to encode, chosen with selectors such as all_nominal_predictors(). |
one_hot |
If TRUE, keep a column for every level instead of dropping a reference. Default FALSE. |
naming |
Function that builds the new column names. Default joins variable and level with an underscore. |
levels |
Filled in by prep(); stores the factor levels learned from the training data. |
skip |
If TRUE, the step is ignored when baking new data. Leave FALSE for encoding. |
Creating dummy variables: worked examples
Build the recipe, prep it, then bake. A recipe is only a plan until prep() records the levels from data. This example dummy-encodes the Species factor in iris using the default treatment contrasts.
Species had three levels, so two dummy columns appear. Row 1 is setosa, the dropped reference, and shows 0 in both columns. The new columns are named variable_level, joining the original name and the factor level with an underscore.
To keep a column for every level, set one_hot = TRUE. This is the encoding tree-free models and neural networks usually expect.
To encode several factors at once, use all_nominal_predictors(). It selects every factor and character predictor while leaving the outcome alone.
Both factors expand to K-1 columns and the numeric hp passes through unchanged. The X prefix appears because level names that start with a digit are not valid R names, so step_dummy() repairs them. To see which levels each factor kept, call tidy() on the prepped recipe with the step number.
step_dummy() vs model.matrix() vs dummyVars()
Pick the encoder that matches your workflow. Base R and caret both create dummy variables, but they differ in how they handle the reference level and new data.
| Approach | Drops a reference level? | Where it fits |
|---|---|---|
step_dummy() |
Yes by default; one_hot = TRUE keeps all |
tidymodels recipes |
model.matrix() |
Yes, always | base R model formulas |
caret::dummyVars() |
No by default | caret preprocessing |
Use step_dummy() when the encoding is part of a tidymodels pipeline, because it stores the factor levels and reapplies them to new data without extra code. Reach for model.matrix() only for a one-off design matrix outside a workflow. caret::dummyVars() is the right call when the surrounding model is already built with caret.
step_other() ahead of step_dummy() to fold infrequent levels into a single other category first.Common pitfalls with step_dummy()
Most failures come from the column type or from unseen levels. Watch these three traps when adding the step to a recipe.
- Passing a numeric or character column you forgot to convert.
step_dummy()expects a factor. A plain character column is coerced, but a numeric column triggers an error, so convert categorical numbers withfactor()first or usestep_num2factor(). - Unseen levels in new data. If
bake()meets a factor level that was absent duringprep(), the dummy columns for that row come back asNA. Addstep_novel()beforestep_dummy()so unknown levels land in a dedicated column instead. - Encoding the outcome.
all_nominal()includes a categorical response variable. Useall_nominal_predictors(), or subtract the outcome with-all_outcomes(), so the target stays a factor for classification.
step_dummy() applies polynomial contrasts by default, producing linear and quadratic terms rather than plain indicators. Convert to an unordered factor first if you want simple 0/1 dummy variables.Try it yourself
Try it: Convert cyl in mtcars to a factor and one-hot encode it in a recipe. Save the baked result to ex_dummies.
Click to reveal solution
Explanation: Setting one_hot = TRUE keeps an indicator column for every level, so all three cylinder counts get a column instead of dropping one as a reference. The X prefix is added because the level names begin with digits.
Related recipes steps
step_dummy() is one of several recipes steps for categorical data. These pair naturally with it in a tidymodels workflow:
- step_other() folds rare factor levels into a single
othergroup. - step_novel() reserves a level for values unseen during training.
- step_unknown() assigns missing factor values to their own level.
- step_interact() builds interaction terms from the dummy columns.
- step_zv() drops indicator columns that hold a single value.
step_dummy() is pd.get_dummies(df), or scikit-learn's OneHotEncoder(). The recipes version differs by learning the factor levels on training data and reapplying the exact same columns to new data automatically.FAQ
What does step_dummy() do in R?
step_dummy() is a recipes step that converts factor or character predictors into numeric 0/1 dummy variables. During prep() it records every level present in the training data. During bake() it creates an indicator column for each level, so a model that needs numeric input can use the categorical information. By default it drops one level as a reference, matching the contrast scheme lm() uses internally.
What is the difference between step_dummy() and one-hot encoding?
Plain step_dummy() uses treatment contrasts: a factor with K levels becomes K-1 columns, with the first level held back as a reference. True one-hot encoding keeps a column for every level. To get one-hot output, set step_dummy(rec, cols, one_hot = TRUE). Use K-1 dummies for linear and penalized regression to avoid collinearity, and one-hot for tree ensembles or neural networks.
Why are my dummy column names prefixed with X?
step_dummy() builds new column names by joining the variable name and the level with an underscore. When a level name starts with a digit, such as 4 or 6, that is not a syntactically valid R name, so step_dummy() repairs it by adding an X prefix. A cyl factor with level 6 becomes the column cyl_X6. You can override this by supplying a custom function to the naming argument.
Do I need to prep() before step_dummy() works?
Yes. A recipe with step_dummy() is just a plan until prep() runs. prep() scans the training data, records the factor levels, and stores them in the step. Only then can bake() produce the dummy columns. Calling bake() on an unprepped recipe throws an error because the levels have not been learned yet.