1 Tidymodels

By now, we have looked at broom (tidy, glance, augment), rsample (splitting data), parsnip (specifying models), and yardstick (evaluating models).

1.1 Workflows

We are now going to look at starting to put together all the pieces into a coherent workflow, similar to how you might chain mutate/group by/summarize/left join/etc in tidyverse as one step.

We are continuing to build our modeling pipeline.

Figure 1.3:

Now let’s introduce the model workflow. We won’t really see the benefit of using this until we start using a lot of different models.

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

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

data(ames)

ames <- ames %>% mutate(Sale_Price = log10(Sale_Price))

set.seed(123)
ames_split <- initial_split(ames, prop = 0.80, strata = Sale_Price)
ames_train <- training(ames_split)
ames_test  <-  testing(ames_split)

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

lm_wflow <- 
  workflow() %>% 
  add_model(lm_model) %>% 
  add_formula(Sale_Price ~ Longitude + Latitude)

lm_fit <- fit(lm_wflow, ames_train)

1.2 Recipies

A recipe is also an object that defines a series of steps for data processing. Unlike the formula method inside a modeling function, the recipe defines the steps without immediately executing them; it is only a specification of what should be done. ~ Sec 8.1

Here’s the example from the textbook:

simple_ames <- 
  recipe(Sale_Price ~ Neighborhood + Gr_Liv_Area + Year_Built + Bldg_Type,
         data = ames_train) %>% #data is used here for column types only
  step_log(Gr_Liv_Area, base = 10) %>% 
  step_dummy(all_nominal_predictors())
simple_ames
## Recipe
## 
## Inputs:
## 
##       role #variables
##    outcome          1
##  predictor          4
## 
## Operations:
## 
## Log transformation on Gr_Liv_Area
## Dummy variables from all_nominal_predictors()

Another key point from 8.3, all preprocessing/feature engineering steps only use the training data. Otherwise, your model might utilize information leakage and lead to overfitting.

There’s a ton of different recipe steps

Let’s go back and look at the divvy data and see how we could use a recipe for data processing. Recall the example from class…

Key steps:

  1. Process recipe using training data
  2. Apply recipe using training data
  3. Apply recipe to testing
divvy_data <- read_csv('https://github.com/erhla/pa470spring2022/raw/main/static/lectures/week_3_data.csv')

grouped <- rsample::initial_split(divvy_data)
train <- training(grouped)
test  <- testing(grouped)


lm_model <-
  parsnip::linear_reg() %>%
  set_engine("lm") %>%
  set_mode('regression')

divvy_rec <- recipe(rides ~ ., data=train)

divvy_rec
## Recipe
## 
## Inputs:
## 
##       role #variables
##    outcome          1
##  predictor          7
divvy_rec <- recipe(rides ~ solar_rad + started_hour +
           temp + wind + interval_rain + avg_speed, data=train) %>%
  step_mutate(hour = factor(hour(started_hour))) %>%
  step_date(started_hour, features=c("dow", "month")) %>%
  step_holiday(started_hour, holidays = timeDate::listHolidays("US"),
               keep_original_cols = FALSE) %>%
  step_dummy(all_nominal_predictors())


#divvy_rec %>% prep() %>% bake(test) %>% view()

divvy_workflow <-
  workflow() %>%
  add_model(lm_model) %>%
  add_recipe(divvy_rec)

divvy_workflow
## ══ Workflow ════════════════════════════════════════════════════════════════════
## Preprocessor: Recipe
## Model: linear_reg()
## 
## ── Preprocessor ────────────────────────────────────────────────────────────────
## 4 Recipe Steps
## 
## • step_mutate()
## • step_date()
## • step_holiday()
## • step_dummy()
## 
## ── Model ───────────────────────────────────────────────────────────────────────
## Linear Regression Model Specification (regression)
## 
## Computational engine: lm
divvy_fit <- divvy_workflow %>%
  fit(data=train)

divvy_fit %>% tidy() %>% view()

divvy_preds <- 
  augment(divvy_fit, test)

Recall that our simple model had a MAPE of 136 and a RMSE of 352.

solar_rad + factor(hour(started_hour)) + 
           factor(wday(started_hour)) +
           factor(month(started_hour)) +
           temp + wind + interval_rain + avg_speed

This is essentially the same model

divvy_fit %>% extract_fit_parsnip() %>% tidy()
yardstick::mape(divvy_preds, 
     truth = rides,
     estimate = .pred)
yardstick::rmse(divvy_preds, 
     truth = rides,
     estimate = .pred)
ggplot(divvy_preds, aes(x=.pred)) +
  geom_density()

Let’s try to improve the model now

divvy_rec <- recipe(rides ~ ., data=train) %>%
  step_mutate(hour = factor(hour(started_hour)),
              bad_weather = if_else(solar_rad <= 5 & temp <= 5, 1, 0),
              nice_weather = if_else(solar_rad >= 25 & temp >= 15, 1, 0)) %>%
  step_date(started_hour, features=c("dow", "month")) %>%
  step_holiday(started_hour, holidays = timeDate::listHolidays("US"),
               keep_original_cols = FALSE) %>%
  step_unknown(all_nominal_predictors()) %>%
  step_dummy(all_nominal_predictors()) %>%
  step_corr(all_predictors())

divvy_workflow <-
  workflow() %>%
  add_model(lm_model) %>%
  add_recipe(divvy_rec)

divvy_workflow
## ══ Workflow ════════════════════════════════════════════════════════════════════
## Preprocessor: Recipe
## Model: linear_reg()
## 
## ── Preprocessor ────────────────────────────────────────────────────────────────
## 6 Recipe Steps
## 
## • step_mutate()
## • step_date()
## • step_holiday()
## • step_unknown()
## • step_dummy()
## • step_corr()
## 
## ── Model ───────────────────────────────────────────────────────────────────────
## Linear Regression Model Specification (regression)
## 
## Computational engine: lm
divvy_fit <- divvy_workflow %>%
  fit(data=train)

divvy_preds <- 
  augment(divvy_fit, test, new_data =  test)


yardstick::mape(divvy_preds, 
     truth = rides,
     estimate = .pred)
yardstick::rmse(divvy_preds, 
     truth = rides,
     estimate = .pred)
ggplot(divvy_preds, aes(x=.pred)) +
  geom_density()

Using decision trees

tree_model <-
  parsnip::decision_tree(tree_depth=5) %>%
  set_engine("rpart") %>%
  set_mode('regression')

divvy_workflow <-
  workflow() %>%
  add_model(tree_model) %>%
  add_recipe(divvy_rec)

divvy_workflow
## ══ Workflow ════════════════════════════════════════════════════════════════════
## Preprocessor: Recipe
## Model: decision_tree()
## 
## ── Preprocessor ────────────────────────────────────────────────────────────────
## 6 Recipe Steps
## 
## • step_mutate()
## • step_date()
## • step_holiday()
## • step_unknown()
## • step_dummy()
## • step_corr()
## 
## ── Model ───────────────────────────────────────────────────────────────────────
## Decision Tree Model Specification (regression)
## 
## Main Arguments:
##   tree_depth = 5
## 
## Computational engine: rpart
divvy_fit <- divvy_workflow %>%
  fit(data=train)

divvy_preds <- 
  augment(divvy_fit, test)


yardstick::mape(divvy_preds, 
     truth = rides,
     estimate = .pred)
yardstick::rmse(divvy_preds, 
     truth = rides,
     estimate = .pred)
ggplot(divvy_preds, aes(x=.pred)) +
  geom_density()

tree_fit <- divvy_fit %>% 
  extract_fit_engine()
rpart.plot::rpart.plot(tree_fit)