Week 5 - Feature Engineering

Download template here

library(tidymodels)
library(nycflights13)

We will not be using any new packages to us today. If the system prompts you to install a package, or gives you a “package not found” error, simply run install.packages("packagename") once to install it.

The data set

We will be using the same flights data set from the nycflights13 package. nycflights13 is an R data package containing all out-bound flights from NYC.

flights

We start by creating the proper response variable before creating the validation split

flights1 <- flights %>%
  mutate(delay = factor(arr_delay > 0, c(TRUE, FALSE),
                        c("Delayed", "On time"))) %>%
  select(-c(year, month, day, hour, minute))

Then we create the test and training data sets.

set.seed(1234)
flights_split <- initial_split(flights1)

flights_train <- training(flights_split)
flights_test <- testing(flights_split)

We will now try some different preprocessing steps with {recipes} to get a hold of it. Every recipe you create starts with a call to recipe.

recipe(delay ~ ., data = flights_train)

then we want to add some steps. Before we add steps, let us think about the order they should be computed in. recipes has a vignette on the Ordering of Steps that gives a reasonnable starting point for ordering.

We saw that we have some missing data. So let us try to rectify that. We have two different types of data here. numeric and nominal data. Both these types have different ways of handing missing data. We will use a simple mean imputation for numerics and step_unknown() to change NAs in the categorical predictors to "unknown"

rec1 <- recipe(delay ~ ., data = flights_train) %>%
  step_impute_mean(all_numeric_predictors()) %>%
  step_unknown(all_nominal_predictors())

We now need to actually calculate the means. We use the prep() function to do that.

rec1_prepped <- rec1 %>%
  prep()
rec1_prepped

We can now apply this recipe to new data. Using the bake() function os what we use here. This function works a lot like predict() in the sense that it takes int a trained/fitted object and is applied to new data.

rec1_prepped %>%
  bake(new_data = flights_train)

We could also in this case have set new_data = NULL since we are using the same data as we supplied to recipe()

rec1_prepped %>%
  bake(new_data = NULL)

We can also investigate the values that were estimated using the tidy() function. Just using tidy() gives us a list of operations that were performed.

rec1_prepped %>%
  tidy()

We can then dig deeper into each operation by specifying the number of the operation. For step_impute_mean() we get the estimated means

rec1_prepped %>%
  tidy(1)

For step_unknown() we see which variables were included and what value get got.

rec1_prepped %>%
  tidy(2)

Next up us dealing with the datetime variable. We will see what happens when we use the features, abbr, label, ordinal and keep_original_cols argument.

rec2 <- recipe(~ time_hour, flights_train) %>%
  step_date(time_hour)

prep(rec2) %>%
  bake(new_data = NULL)

YOUR TURN

Create a single recipe, containing the above steps. Create dummy variables of all the nominal predictors with step_dummy() and end with a normalization step via step_normalize(). Prep it on the training data set, and bake it on the testing data set.