In workflows: Modeling Workflows

Formula Handling

options(width = 70)

library(parsnip)
library(workflows)
library(magrittr)
library(modeldata)
library(hardhat)
library(splines)

Note that, for different models, the formula given to add_formula() might be handled in different ways, depending on the parsnip model being used. For example, a random forest model fit using ranger would not convert any factor predictors to binary indicator variables. This is consistent with what ranger::ranger() would do, but is inconsistent with what stats::model.matrix() would do.

The documentation for parsnip models provides details about how the data given in the formula are encoded for the model if they diverge from the standard model.matrix() methodology. Our goal is to be consistent with how the underlying model package works.

How is this formula used?

To demonstrate, the example below uses lm() to fit a model. The formula given to add_formula() is used to create the model matrix and that is what is passed to lm() with a simple formula of body_mass_g ~ .:

library(parsnip)
library(workflows)
library(magrittr)
library(modeldata)
library(hardhat)

data(penguins)

lm_mod <- linear_reg() %>% 
  set_engine("lm")

lm_wflow <- workflow() %>% 
  add_model(lm_mod)

pre_encoded <- lm_wflow %>% 
  add_formula(body_mass_g ~ species + island + bill_depth_mm) %>% 
  fit(data = penguins)

pre_encoded_parsnip_fit <- pre_encoded %>% 
  extract_fit_parsnip()

pre_encoded_fit <- pre_encoded_parsnip_fit$fit

# The `lm()` formula is *not* the same as the `add_formula()` formula: 
pre_encoded_fit

This can affect how the results are analyzed. For example, to get sequential hypothesis tests, each individual term is tested:

anova(pre_encoded_fit)

Overriding the default encodings

Users can override the model-specific encodings by using a hardhat blueprint. The blueprint can specify how factors are encoded and whether intercepts are included. As an example, if you use a formula and would like the data to be passed to a model untouched:

minimal <- default_formula_blueprint(indicators = "none", intercept = FALSE)

un_encoded <- lm_wflow %>% 
  add_formula(
    body_mass_g ~ species + island + bill_depth_mm, 
    blueprint = minimal
  ) %>% 
  fit(data = penguins)

un_encoded_parsnip_fit <- un_encoded %>% 
  extract_fit_parsnip()

un_encoded_fit <- un_encoded_parsnip_fit$fit

un_encoded_fit

While this looks the same, the raw columns were given to lm() and that function created the dummy variables. Because of this, the sequential ANOVA tests groups of parameters to get column-level p-values:

anova(un_encoded_fit)

Overriding the default model formula

Additionally, the formula passed to the underlying model can also be customized. In this case, the formula argument of add_model() can be used. To demonstrate, a spline function will be used for the bill depth:

library(splines)

custom_formula <- workflow() %>%
  add_model(
    lm_mod, 
    formula = body_mass_g ~ species + island + ns(bill_depth_mm, 3)
  ) %>% 
  add_formula(
    body_mass_g ~ species + island + bill_depth_mm, 
    blueprint = minimal
  ) %>% 
  fit(data = penguins)

custom_parsnip_fit <- custom_formula %>% 
  extract_fit_parsnip()

custom_fit <- custom_parsnip_fit$fit

custom_fit

Altering the formula

Finally, when a formula is updated or removed from a fitted workflow, the corresponding model fit is removed.

custom_formula_no_fit <- update_formula(custom_formula, body_mass_g ~ species)

try(extract_fit_parsnip(custom_formula_no_fit))

workflows documentation built on May 29, 2024, 3:57 a.m.