Fitting a workflow currently involves two main steps:
Preprocessing the data using a formula preprocessor, or by calling
recipes::prep() on a recipe.
Fitting the underlying parsnip model using parsnip::fit.model_spec().
# S3 method for workflow
fit(object, data, ..., control = control_workflow())A workflow
A data frame of predictors and outcomes to use when fitting the workflow
Not used
A control_workflow() object
The workflow object, updated with a fit parsnip model in the
object$fit$fit slot.
Some modeling functions in R create indicator/dummy variables from
categorical data when you use a model formula, and some do not. When you
specify and fit a model with a workflow(), parsnip and workflows match
and reproduce the underlying behavior of the user-specified model<U+2019>s
computational engine.
In the modeldata::Sacramento data set of real
estate prices, the type variable has three levels: "Residential",
"Condo", and "Multi-Family". This base workflow() contains a
formula added via add_formula() to predict
property price from property type, square footage, number of beds, and
number of baths:
set.seed(123)
library(parsnip) library(recipes) library(workflows) library(modeldata)
data("Sacramento")
base_wf <- workflow() %>% add_formula(price ~ type + sqft + beds + baths)
This first model does create dummy/indicator variables:
lm_spec <- linear_reg() %>%
set_engine("lm")base_wf %>% add_model(lm_spec) %>% fit(Sacramento)
## == Workflow [trained] ================================================ ## Preprocessor: Formula ## Model: linear_reg() ## ## -- Preprocessor ------------------------------------------------------ ## price ~ type + sqft + beds + baths ## ## -- Model ------------------------------------------------------------- ## ## Call: ## stats::lm(formula = ..y ~ ., data = data) ## ## Coefficients: ## (Intercept) typeMulti_Family typeResidential ## 32919.4 -21995.8 33688.6 ## sqft beds baths ## 156.2 -29788.0 8730.0
There are five independent variables in the fitted model for this
OLS linear regression. With this model type and engine, the factor
predictor type of the real estate properties was converted to two
binary predictors, typeMulti_Family and typeResidential. (The third
type, for condos, does not need its own column because it is the
baseline level).
This second model does not create dummy/indicator variables:
rf_spec <- rand_forest() %>%
set_mode("regression") %>%
set_engine("ranger")base_wf %>% add_model(rf_spec) %>% fit(Sacramento)
## == Workflow [trained] ================================================ ## Preprocessor: Formula ## Model: rand_forest() ## ## -- Preprocessor ------------------------------------------------------ ## price ~ type + sqft + beds + baths ## ## -- Model ------------------------------------------------------------- ## Ranger result ## ## Call: ## ranger::ranger(formula = ..y ~ ., data = data, num.threads = 1, verbose = FALSE, seed = sample.int(10^5, 1)) ## ## Type: Regression ## Number of trees: 500 ## Sample size: 932 ## Number of independent variables: 4 ## Mtry: 2 ## Target node size: 5 ## Variable importance mode: none ## Splitrule: variance ## OOB prediction error (MSE): 7058847504 ## R squared (OOB): 0.5894647
Note that there are four independent variables in the fitted model
for this ranger random forest. With this model type and engine,
indicator variables were not created for the type of real estate
property being sold. Tree-based models such as random forest models can
handle factor predictors directly, and don<U+2019>t need any conversion to
numeric binary variables.
When you specify a model with a workflow() and a recipe preprocessor
via add_recipe(), the recipe controls whether
dummy variables are created or not; the recipe overrides any underlying
behavior from the model<U+2019>s computational engine.
In the future, there will also be postprocessing steps that can be added after the model has been fit.
# NOT RUN {
library(parsnip)
library(recipes)
library(magrittr)
model <- linear_reg() %>%
set_engine("lm")
base_wf <- workflow() %>%
add_model(model)
formula_wf <- base_wf %>%
add_formula(mpg ~ cyl + log(disp))
fit(formula_wf, mtcars)
recipe <- recipe(mpg ~ cyl + disp, mtcars) %>%
step_log(disp)
recipe_wf <- base_wf %>%
add_recipe(recipe)
fit(recipe_wf, mtcars)
# }
Run the code above in your browser using DataCamp Workspace