In our Build a Model article, we learned how to specify and train models with different engines using the parsnip package. In this article, we’ll explore another tidymodels package, recipes, which is designed to help you preprocess your data before training your model. Recipes are built as a series of preprocessing steps, such as:
converting qualitative predictors to indicator variables (also known as dummy variables),
transforming data to be on a different scale (e.g., taking the logarithm of a variable),
transforming whole groups of predictors together,
extracting key features from raw variables (e.g., getting the day of the week out of a date variable),
and so on. If you are familiar with R’s formula interface, a lot of this might sound familiar and like what a formula already does. Recipes can be used to do many of the same things, but they have a much wider range of possibilities. This article shows how to use recipes for modeling.
To use code in this article, you will need to install the following packages: nycflights13, skimr, and tidymodels.
library(tidymodels) # for the recipes package, along with the rest of tidymodels # Helper packages library(nycflights13) # for flight data library(skimr) # for variable summaries
Let’s use the nycflights13 data to predict whether a plane arrives more than 30 minutes late. This data set contains information on 325,819 flights departing near New York City in 2013. Let’s start by loading the data and making a few changes to the variables:
set.seed(123) flight_data <- flights %>% mutate( # Convert the arrival delay to a factor arr_delay = ifelse(arr_delay >= 30, "late", "on_time"), arr_delay = factor(arr_delay), # We will use the date (not date-time) in the recipe below date = as.Date(time_hour) ) %>% # Include the weather data inner_join(weather, by = c("origin", "time_hour")) %>% # Only retain the specific columns we will use select(dep_time, flight, origin, dest, air_time, distance, carrier, date, arr_delay, time_hour) %>% # Exclude missing data na.omit() %>% # For creating models, it is better to have qualitative columns # encoded as factors (instead of character strings) mutate_if(is.character, as.factor)
We can see that about 16% of the flights in this data set arrived more than 30 minutes late.
flight_data %>% count(arr_delay) %>% mutate(prop = n/sum(n)) #> # A tibble: 2 x 3 #> arr_delay n prop #> <fct> <int> <dbl> #> 1 late 52540 0.161 #> 2 on_time 273279 0.839
Before we start building up our recipe, let’s take a quick look at a few specific variables that will be important for both preprocessing and modeling.
First, notice that the variable we created called
arr_delay is a factor variable; it is important that our outcome variable for training a logistic regression model is a factor.
glimpse(flight_data) #> Observations: 325,819 #> Variables: 10 #> $ dep_time <int> 517, 533, 542, 544, 554, 554, 555, 557, 557, 558, 558, 558,… #> $ flight <int> 1545, 1714, 1141, 725, 461, 1696, 507, 5708, 79, 301, 49, 7… #> $ origin <fct> EWR, LGA, JFK, JFK, LGA, EWR, EWR, LGA, JFK, LGA, JFK, JFK,… #> $ dest <fct> IAH, IAH, MIA, BQN, ATL, ORD, FLL, IAD, MCO, ORD, PBI, TPA,… #> $ air_time <dbl> 227, 227, 160, 183, 116, 150, 158, 53, 140, 138, 149, 158, … #> $ distance <dbl> 1400, 1416, 1089, 1576, 762, 719, 1065, 229, 944, 733, 1028… #> $ carrier <fct> UA, UA, AA, B6, DL, UA, B6, EV, B6, AA, B6, B6, UA, UA, AA,… #> $ date <date> 2013-01-01, 2013-01-01, 2013-01-01, 2013-01-01, 2013-01-01… #> $ arr_delay <fct> on_time, on_time, late, on_time, on_time, on_time, on_time,… #> $ time_hour <dttm> 2013-01-01 05:00:00, 2013-01-01 05:00:00, 2013-01-01 05:00…
Second, there are two variables that we don’t want to use as predictors in our model, but that we would like to retain as identification variables that can be used to troubleshoot poorly predicted data points. These are
flight, a numeric value, and
time_hour, a date-time value.
Third, there are 104 flight destinations contained in
dest and 16 distinct
flight_data %>% skimr::skim(dest, carrier)
|Number of rows||325819|
|Number of columns||10|
|Column type frequency:|
Variable type: factor
|dest||0||1||FALSE||104||ATL: 16771, ORD: 16507, LAX: 15942, BOS: 14948|
|carrier||0||1||FALSE||16||UA: 57489, B6: 53715, EV: 50868, DL: 47465|
Because we’ll be using a simple logistic regression model, the variables
carrier will be converted to
dummy variables. However, some of these values do not occur very frequently and this could complicate our analysis. We’ll discuss specific steps later in this article that we can add to our recipe to address this issue before modeling.
To get started, let’s split this single dataset into two: a training set and a testing set. We’ll keep most of the rows in the original dataset (subset chosen randomly) in the training set. The training data will be used to fit the model, and the testing set will be used to measure model performance.
To do this, we can use the rsample package to create an object that contains the information on how to split the data, and then two more rsample functions to create data frames for the training and testing sets:
# Fix the random numbers by setting the seed # This enables the analysis to be reproducible when random numbers are used set.seed(555) # Put 3/4 of the data into the training set data_split <- initial_split(flight_data, prop = 3/4) # Create data frames for the two sets: train_data <- training(data_split) test_data <- testing(data_split)
To get started, let’s create a recipe for a simple logistic regression model. Before training the model, we can use a recipe to create a few new predictors and conduct some preprocessing required by the model.
Let’s initiate a new recipe:
flights_rec <- recipe(arr_delay ~ ., data = train_data)
recipe() function as we used it here has two arguments:
A formula. Any variable on the left-hand side of the tilde (
~) is considered the model outcome (here,
arr_delay). On the right-hand side of the tilde are the predictors. Variables may be listed by name, or you can use the dot (
.) to indicate all other variables as predictors.
The data. A recipe is associated with the data set used to create the model. This will typically be the training set, so
data = train_datahere. Naming a data set doesn’t actually change the data itself; it is only used to catalog the names of the variables and their types, like factors, integers, dates, etc.
Now we can add
roles to this recipe. We can use the
update_role() function to let recipes know that
time_hour are variables with a custom role that we called
"ID" (a role can have any character value). Whereas our formula included all variables in the training set other than
arr_delay as predictors, this tells the recipe to keep these two variables but not use them as either outcomes or predictors.
flights_rec <- recipe(arr_delay ~ ., data = train_data) %>% update_role(flight, time_hour, new_role = "ID")
This step of adding roles to a recipe is optional; the purpose of using it here is that those two variables can be retained in the data but not included in the model. This can be convenient when, after the model is fit, we want to investigate some poorly predicted value. These ID columns will be available and can be used to try to understand what went wrong.
To get the current set of variables and roles, use the
summary(flights_rec) #> # A tibble: 10 x 4 #> variable type role source #> <chr> <chr> <chr> <chr> #> 1 dep_time numeric predictor original #> 2 flight numeric ID original #> 3 origin nominal predictor original #> 4 dest nominal predictor original #> 5 air_time numeric predictor original #> 6 distance numeric predictor original #> 7 carrier nominal predictor original #> 8 date date predictor original #> 9 time_hour date ID original #> 10 arr_delay nominal outcome original
Now we can start adding steps onto our recipe using the pipe operator. Perhaps it is reasonable for the date of the flight to have an effect on the likelihood of a late arrival. A little bit of feature engineering might go a long way to improving our model. How should the date be encoded into the model? The
date column has an R
date object so including that column “as is” will mean that the model will convert it to a numeric format equal to the number of days after a reference date:
flight_data %>% distinct(date) %>% mutate(numeric_date = as.numeric(date)) #> # A tibble: 364 x 2 #> date numeric_date #> <date> <dbl> #> 1 2013-01-01 15706 #> 2 2013-01-02 15707 #> 3 2013-01-03 15708 #> 4 2013-01-04 15709 #> 5 2013-01-05 15710 #> # … with 359 more rows
It’s possible that the numeric date variable is a good option for modeling; perhaps the model would benefit from a linear trend between the log-odds of a late arrival and the numeric date variable. However, it might be better to add model terms derived from the date that have a better potential to be important to the model. For example, we could derive the following meaningful features from the single
the day of the week,
the month, and
whether or not the date corresponds to a holiday.
Let’s do all three of these by adding steps to our recipe:
flights_rec <- recipe(arr_delay ~ ., data = train_data) %>% update_role(flight, time_hour, new_role = "ID") %>% step_date(date, features = c("dow", "month")) %>% step_holiday(date, holidays = timeDate::listHolidays("US")) %>% step_rm(date)
What do each of these steps do?
step_date(), we created two new factor columns with the appropriate day of the week and the month.
step_holiday(), we created a binary variable indicating whether the current date is a holiday or not. The argument value of
timeDate::listHolidays("US")uses the timeDate package to list the 17 standard US holidays.
step_rm(), we remove the original
datevariable since we no longer want it in the model.
Next, we’ll turn our attention to the variable types of our predictors. Because we plan to train a logistic regression model, we know that predictors will ultimately need to be numeric, as opposed to factor variables. In other words, there may be a difference in how we store our data (in factors inside a data frame), and how the underlying equations require them (a purely numeric matrix).
For factors like
standard practice is to convert them into dummy or indicator variables to make them numeric. These are binary values for each level of the factor. For example, our
origin variable has values of
"LGA". The standard dummy variable encoding, shown below, will create two numeric columns of the data that are 1 when the originating airport is
"LGA" and zero otherwise, respectively.
But, unlike the standard model formula methods in R, a recipe does not automatically create these dummy variables for you; you’ll need to tell your recipe to add this step. This is for two reasons. First, many models do not require
numeric predictors, so dummy variables may not always be preferred. Second, recipes can also be used for purposes outside of modeling, where non-dummy versions of the variables may work better. For example, you may want to make a table or a plot with a variable as a single factor. For those reasons, you need to explicitly tell recipes to create dummy variables using
flights_rec <- recipe(arr_delay ~ ., data = train_data) %>% update_role(flight, time_hour, new_role = "ID") %>% step_date(date, features = c("dow", "month")) %>% step_holiday(date, holidays = timeDate::listHolidays("US")) %>% step_rm(date) %>% step_dummy(all_nominal(), -all_outcomes())
Here, we did something different than before: instead of applying a step to an individual variable, we used selectors to apply this recipe step to several variables at once.
The first selector,
all_nominal(), selects all variables that are either factors or characters.
The second selector,
-all_outcomes()removes any outcome variables from this recipe step.
With these two selectors together, our recipe step above translates to:
Create dummy variables for all of the factor or character columns unless they are outcomes.
At this stage in the recipe, this step selects the
carrier variables. It also includes two new variables,
date_month, that were created by the earlier
More generally, the recipe selectors mean that you don’t always have to apply steps to individual variables one at a time. Since a recipe knows the variable type and role of each column, they can also be selected (or dropped) using this information.
We need one final step to add to our recipe. Since
dest have some infrequently occurring values, it is possible that dummy variables might be created for values that don’t exist in the training set. For example, there is one destination that is only in the test set:
test_data %>% distinct(dest) %>% anti_join(train_data) #> Joining, by = "dest" #> # A tibble: 1 x 1 #> dest #> <fct> #> 1 LEX
When the recipe is applied to the training set, a column is made for LEX but it will contain all zeros. This is a “zero-variance predictor” that has no information within the column. While some R functions will not produce an error for such predictors, it usually causes warnings and other issues.
step_zv() will remove columns from the data when the training set data have a single value, so it is added to the recipe after
flights_rec <- recipe(arr_delay ~ ., data = train_data) %>% update_role(flight, time_hour, new_role = "ID") %>% step_date(date, features = c("dow", "month")) %>% step_holiday(date, holidays = timeDate::listHolidays("US")) %>% step_rm(date) %>% step_dummy(all_nominal(), -all_outcomes()) %>% step_zv(all_predictors())
Now we’ve created a specification of what should be done with the data. How do we use the recipe we made?
lr_mod <- logistic_reg() %>% set_engine("glm")
We will want to use our recipe across several steps as we train and test our model. We will:
Process the recipe using the training set: This involves any estimation or calculations based on the training set. For our recipe, the training set will be used to determine which predictors should be converted to dummy variables and which predictors will have zero-variance in the training set, and should be slated for removal.
Apply the recipe to the training set: We create the final predictor set on the training set.
Apply the recipe to the test set: We create the final predictor set on the test set. Nothing is recomputed and no information from the test set is used here; the dummy variable and zero-variance results from the training set are applied to the test set.
To simplify this process, we can use a model workflow, which pairs a model and recipe together. This is a straightforward approach because different recipes are often needed for different models, so when a model and recipe are bundled, it becomes easier to train and test workflows. We’ll use the
workflows package from tidymodels to bundle our parsnip model (
lr_mod) with our recipe (
flights_wflow <- workflow() %>% add_model(lr_mod) %>% add_recipe(flights_rec) flights_wflow #> ══ Workflow ═════════════════════════════════════════════════════════════ #> Preprocessor: Recipe #> Model: logistic_reg() #> #> ── Preprocessor ───────────────────────────────────────────────────────── #> 5 Recipe Steps #> #> ● step_date() #> ● step_holiday() #> ● step_rm() #> ● step_dummy() #> ● step_zv() #> #> ── Model ──────────────────────────────────────────────────────────────── #> Logistic Regression Model Specification (classification) #> #> Computational engine: glm
Now, there is a single function that can be used to prepare the recipe and train the model from the resulting predictors:
flights_fit <- flights_wflow %>% fit(data = train_data)
This object has the finalized recipe and fitted model objects inside. You may want to extract the model or recipe objects from the workflow. To do this, you can use the helper functions
pull_workflow_prepped_recipe(). For example, here we pull the fitted model object then use the
broom::tidy() function to get a tidy tibble of model coefficients:
flights_fit %>% pull_workflow_fit() %>% tidy() #> # A tibble: 157 x 5 #> term estimate std.error statistic p.value #> <chr> <dbl> <dbl> <dbl> <dbl> #> 1 (Intercept) 3.91 2.73 1.43 1.51e- 1 #> 2 dep_time -0.00167 0.0000141 -118. 0. #> 3 air_time -0.0439 0.000561 -78.4 0. #> 4 distance 0.00686 0.00150 4.57 4.84e- 6 #> 5 date_USChristmasDay 1.12 0.173 6.49 8.45e-11 #> # … with 152 more rows
Our goal was to predict whether a plane arrives more than 30 minutes late. We have just:
Built the model (
Created a preprocessing recipe (
Bundled the model and recipe (
Trained our workflow using a single call to
The next step is to use the trained workflow (
flights_fit) to predict with the unseen test data, which we will do with a single call to
predict() method applies the recipe to the new data, then passes them to the fitted model.
predict(flights_fit, test_data) #> # A tibble: 81,454 x 1 #> .pred_class #> <fct> #> 1 on_time #> 2 on_time #> 3 on_time #> 4 on_time #> 5 on_time #> # … with 8.145e+04 more rows
Because our outcome variable here is a factor, the output from
predict() returns the predicted class:
on_time. But, let’s say we want the predicted class probabilities for each flight instead. To return those, we can specify
type = "prob" when we use
predict(). We’ll also bind the output with some variables from the test data and save them together:
flights_pred <- predict(flights_fit, test_data, type = "prob") %>% bind_cols(test_data %>% select(arr_delay, time_hour, flight)) # The data look like: flights_pred #> # A tibble: 81,454 x 5 #> .pred_late .pred_on_time arr_delay time_hour flight #> <dbl> <dbl> <fct> <dttm> <int> #> 1 0.0565 0.944 on_time 2013-01-01 05:00:00 1714 #> 2 0.0264 0.974 on_time 2013-01-01 06:00:00 79 #> 3 0.0481 0.952 on_time 2013-01-01 06:00:00 301 #> 4 0.0325 0.967 on_time 2013-01-01 06:00:00 49 #> 5 0.0711 0.929 on_time 2013-01-01 06:00:00 1187 #> # … with 8.145e+04 more rows
Now that we have a tibble with our predicted class probabilities, how will we evaluate the performance of our workflow? We can see from these first few rows that our model predicted these 5 on time flights correctly because the values of
.pred_on_time are p > .50. But we also know that we have 81,454 rows total to predict. We would like to calculate a metric that tells how well our model predicted late arrivals, compared to the true status of our outcome variable,
To generate a ROC curve, we need the predicted class probabilities for
on_time, which we just calculated in the code chunk above. We can create the ROC curve with these values, using
roc_curve() and then piping to the
flights_pred %>% roc_curve(truth = arr_delay, .pred_late) %>% autoplot()
roc_auc() estimates the area under the curve:
flights_pred %>% roc_auc(truth = arr_delay, .pred_late) #> # A tibble: 1 x 3 #> .metric .estimator .estimate #> <chr> <chr> <dbl> #> 1 roc_auc binary 0.765
Not too bad! We leave it to the reader to test out this workflow
without this recipe. You can use
workflows::add_formula(arr_delay ~ .) instead of
add_recipe() (remember to remove the identification variables first!), and see whether our recipe improved our model’s ability to predict late arrivals.
#> ─ Session info ─────────────────────────────────────────────────────────────── #> setting value #> version R version 3.6.2 (2019-12-12) #> os macOS Mojave 10.14.6 #> system x86_64, darwin15.6.0 #> ui X11 #> language (EN) #> collate en_US.UTF-8 #> ctype en_US.UTF-8 #> tz America/Denver #> date 2020-04-20 #> #> ─ Packages ─────────────────────────────────────────────────────────────────── #> package * version date lib source #> broom * 0.5.5 2020-02-29  CRAN (R 3.6.0) #> dials * 0.0.6 2020-04-03  CRAN (R 3.6.2) #> dplyr * 0.8.5 2020-03-07  CRAN (R 3.6.0) #> ggplot2 * 3.3.0 2020-03-05  CRAN (R 3.6.0) #> infer * 0.5.1 2019-11-19  CRAN (R 3.6.0) #> nycflights13 * 1.0.1 2019-09-16  CRAN (R 3.6.0) #> parsnip * 0.1.0 2020-04-09  CRAN (R 3.6.2) #> purrr * 0.3.4 2020-04-17  CRAN (R 3.6.2) #> recipes * 0.1.10 2020-03-18  CRAN (R 3.6.0) #> rlang 0.4.5 2020-03-01  CRAN (R 3.6.0) #> rsample * 0.0.6 2020-03-31  CRAN (R 3.6.2) #> skimr * 2.1.1 2020-04-16  CRAN (R 3.6.2) #> tibble * 2.1.3 2019-06-06  CRAN (R 3.6.2) #> tidymodels * 0.1.0 2020-02-16  CRAN (R 3.6.0) #> tune * 0.1.0 2020-04-02  CRAN (R 3.6.2) #> workflows * 0.1.1 2020-03-17  CRAN (R 3.6.0) #> yardstick * 0.0.6 2020-03-17  CRAN (R 3.6.0) #> #>  /Library/Frameworks/R.framework/Versions/3.6/Resources/library