1 - Feature Engineering

Advanced tidymodels

Working with our predictors

We might want to modify our predictors columns for a few reasons:

• The model requires them in a different format (e.g. dummy variables for linear regression).
• The model needs certain data qualities (e.g. same units for K-NN).
• The outcome is better predicted when one or more columns are transformed in some way (a.k.a “feature engineering”).

The first two reasons are fairly predictable (next page).

The last one depends on your modeling problem.

What is feature engineering?

Think of a feature as some representation of a predictor that will be used in a model.

Example representations:

• Interactions
• Polynomial expansions/splines
• Principal component analysis (PCA) feature extraction

There are a lot of examples in Feature Engineering and Selection (FES).

Example: Dates

How can we represent date columns for our model?

When we use a date column in its native format, most models in R convert it to an integer.

We can re-engineer it as:

• Days since a reference date
• Day of the week
• Month
• Year
• Indicators for holidays

The main point is that we try to maximize performance with different versions of the predictors.

Mention that, for the Chicago data, the day or the week features are usually the most important ones in the model.

General definitions

• Data preprocessing steps allow your model to fit.

• Feature engineering steps help the model do the least work to predict the outcome as well as possible.

The recipes package can handle both!

These terms are often used interchangeably in the ML community but we want to distinguish them.

Hotel Data

We’ll use data on hotels to predict the cost of a room.

The data are in the modeldatatoo package. We’ll sample down the data and refactor some columns:

library(tidymodels)
library(modeldatatoo)

# Max's usual settings: 
tidymodels_prefer()
theme_set(theme_bw())
options(
  pillar.advice = FALSE, 
  pillar.min_title_chars = Inf
)

set.seed(295)
hotel_rates <- 
  data_hotel_rates() %>% 
  sample_n(5000) %>% 
  arrange(arrival_date) %>% 
  select(-arrival_date_num, -arrival_date) %>% 
  mutate(
    company = factor(as.character(company)),
    country = factor(as.character(country)),
    agent = factor(as.character(agent))
  )

Data splitting strategy

Data Spending

Let’s split the data into a training set (75%) and testing set (25%):

set.seed(4028)
hotel_split <-
  initial_split(hotel_rates, strata = avg_price_per_room)

hotel_tr <- training(hotel_split)
hotel_te <- testing(hotel_split)

Your turn

Let’s take some time and investigate the training data. The outcome is avg_price_per_room.

Are there any interesting characteristics of the data?

10:00

Resampling Strategy

Resampling Strategy

We’ll use simple 10-fold cross-validation (stratified sampling):

set.seed(472)
hotel_rs <- vfold_cv(hotel_tr, strata = avg_price_per_room)
hotel_rs
#> #  10-fold cross-validation using stratification 
#> # A tibble: 10 × 2
#>    splits             id    
#>    <list>             <chr> 
#>  1 <split [3372/377]> Fold01
#>  2 <split [3373/376]> Fold02
#>  3 <split [3373/376]> Fold03
#>  4 <split [3373/376]> Fold04
#>  5 <split [3373/376]> Fold05
#>  6 <split [3374/375]> Fold06
#>  7 <split [3375/374]> Fold07
#>  8 <split [3376/373]> Fold08
#>  9 <split [3376/373]> Fold09
#> 10 <split [3376/373]> Fold10

Prepare your data for modeling

• The recipes package is an extensible framework for pipeable sequences of preprocessing and feature engineering steps.

• Statistical parameters for the steps can be estimated from an initial data set and then applied to other data sets.

• The resulting processed output can be used as inputs for statistical or machine learning models.

A first recipe

hotel_rec <- 
  recipe(avg_price_per_room ~ ., data = hotel_tr)

• The recipe() function assigns columns to roles of “outcome” or “predictor” using the formula

A first recipe

summary(hotel_rec)
#> # A tibble: 28 × 4
#>    variable                  type      role      source  
#>    <chr>                     <list>    <chr>     <chr>   
#>  1 lead_time                 <chr [2]> predictor original
#>  2 arrival_date_day_of_month <chr [2]> predictor original
#>  3 stays_in_weekend_nights   <chr [2]> predictor original
#>  4 stays_in_week_nights      <chr [2]> predictor original
#>  5 adults                    <chr [2]> predictor original
#>  6 children                  <chr [2]> predictor original
#>  7 babies                    <chr [2]> predictor original
#>  8 meal                      <chr [3]> predictor original
#>  9 country                   <chr [3]> predictor original
#> 10 market_segment            <chr [3]> predictor original
#> # ℹ 18 more rows

The type column contains information on the variables

Your turn

What do you think are in the type vectors for the lead_time and country columns?

02:00

Create indicator variables

hotel_rec <- 
  recipe(avg_price_per_room ~ ., data = hotel_tr) %>% 
  step_dummy(all_nominal_predictors())

• For any factor or character predictors, make binary indicators.

• There are many recipe steps that can convert categorical predictors to numeric columns.

• step_dummy() records the levels of the categorical predictors in the training set.

Filter out constant columns

hotel_rec <- 
  recipe(avg_price_per_room ~ ., data = hotel_tr) %>% 
  step_dummy(all_nominal_predictors()) %>% 
  step_zv(all_predictors())

In case there is a factor level that was never observed in the training data (resulting in a column of all 0s), we can delete any zero-variance predictors that have a single unique value.

Note that the selector chooses all columns with a role of “predictor”

Normalization

hotel_rec <- 
  recipe(avg_price_per_room ~ ., data = hotel_tr) %>% 
  step_dummy(all_nominal_predictors()) %>% 
  step_zv(all_predictors()) %>% 
  step_normalize(all_numeric_predictors())

• This centers and scales the numeric predictors.

• The recipe will use the training set to estimate the means and standard deviations of the data.

• All data the recipe is applied to will be normalized using those statistics (there is no re-estimation).

Reduce correlation

hotel_rec <- 
  recipe(avg_price_per_room ~ ., data = hotel_tr) %>% 
  step_dummy(all_nominal_predictors()) %>% 
  step_zv(all_predictors()) %>% 
  step_normalize(all_numeric_predictors()) %>% 
  step_corr(all_numeric_predictors(), threshold = 0.9)

To deal with highly correlated predictors, find the minimum set of predictor columns that make the pairwise correlations less than the threshold.

Other possible steps

hotel_rec <- 
  recipe(avg_price_per_room ~ ., data = hotel_tr) %>% 
  step_dummy(all_nominal_predictors()) %>% 
  step_zv(all_predictors()) %>% 
  step_normalize(all_numeric_predictors()) %>% 
  step_pca(all_numeric_predictors())

PCA feature extraction…

Other possible steps

hotel_rec <- 
  recipe(avg_price_per_room ~ ., data = hotel_tr) %>% 
  step_dummy(all_nominal_predictors()) %>% 
  step_zv(all_predictors()) %>% 
  step_normalize(all_numeric_predictors()) %>% 
  embed::step_umap(all_numeric_predictors(), outcome = avg_price_per_room)

A fancy machine learning supervised dimension reduction technique…

Note that this uses the outcome, and it is from an extension package

Other possible steps

hotel_rec <- 
  recipe(avg_price_per_room ~ ., data = hotel_tr) %>% 
  step_dummy(all_nominal_predictors()) %>% 
  step_zv(all_predictors()) %>% 
  step_normalize(all_numeric_predictors()) %>% 
  step_spline_natural(year_day, deg_free = 10)

Nonlinear transforms like natural splines, and so on!

Your turn

Create a recipe() for the hotel data to:

• use a Yeo-Johnson (YJ) transformation on lead_time
• convert factors to indicator variables
• remove zero-variance variables

03:00

Minimal recipe

hotel_indicators <-
  recipe(avg_price_per_room ~ ., data = hotel_tr) %>% 
  step_YeoJohnson(lead_time) %>%
  step_dummy(all_nominal_predictors()) %>%
  step_zv(all_predictors())

Measuring Performance

We’ll compute two measures: mean absolute error and the coefficient of determination (a.k.a \(R^2\)).

\[\begin{align} MAE &= \frac{1}{n}\sum_{i=1}^n |y_i - \hat{y}_i| \notag \\ R^2 &= cor(y_i, \hat{y}_i)^2 \end{align}\]

The focus will be on MAE for parameter optimization. We’ll use a metric set to compute these:

reg_metrics <- metric_set(mae, rsq)

Using a workflow

set.seed(9)

hotel_lm_wflow <-
  workflow() %>%
  add_recipe(hotel_indicators) %>%
  add_model(linear_reg())
 
ctrl <- control_resamples(save_pred = TRUE)
hotel_lm_res <-
  hotel_lm_wflow %>%
  fit_resamples(hotel_rs, control = ctrl, metrics = reg_metrics)

collect_metrics(hotel_lm_res)
#> # A tibble: 2 × 6
#>   .metric .estimator   mean     n std_err .config             
#>   <chr>   <chr>       <dbl> <int>   <dbl> <chr>               
#> 1 mae     standard   17.3      10 0.199   Preprocessor1_Model1
#> 2 rsq     standard    0.874    10 0.00400 Preprocessor1_Model1

Your turn

Use fit_resamples() to fit your workflow with a recipe.

Collect the predictions from the results.

05:00

Holdout predictions

# Since we used `save_pred = TRUE`
lm_val_pred <- collect_predictions(hotel_lm_res)
lm_val_pred %>% slice(1:7)
#> # A tibble: 7 × 5
#>   id     .pred  .row avg_price_per_room .config             
#>   <chr>  <dbl> <int>              <dbl> <chr>               
#> 1 Fold01  62.1    20                 40 Preprocessor1_Model1
#> 2 Fold01  48.0    28                 54 Preprocessor1_Model1
#> 3 Fold01  64.6    45                 50 Preprocessor1_Model1
#> 4 Fold01  45.8    49                 42 Preprocessor1_Model1
#> 5 Fold01  45.8    61                 49 Preprocessor1_Model1
#> 6 Fold01  30.0    66                 40 Preprocessor1_Model1
#> 7 Fold01  38.8    88                 49 Preprocessor1_Model1

Calibration Plot

library(probably)

cal_plot_regression(hotel_lm_res, alpha = 1 / 5)

What do we do with the agent and company data?

There are 98 unique agent values and 100 unique companies in our training set. How can we include this information in our model?

We could:

• make the full set of indicator variables 😳

• lump agents and companies that rarely occur into an “other” group

• use feature hashing to create a smaller set of indicator variables

• use effect encoding to replace the agent and company columns with the estimated effect of that predictor (in the extra materials)

Per-agent statistics

Collapsing factor levels

There is a recipe step that will redefine factor levels based on their frequency in the training set:

hotel_other_rec <-
  recipe(avg_price_per_room ~ ., data = hotel_tr) %>% 
  step_YeoJohnson(lead_time) %>%
  step_other(agent, threshold = 0.001) %>%
  step_dummy(all_nominal_predictors()) %>%
  step_zv(all_predictors())

Using this code, 34 agents (out of 98) were collapsed into “other” based on the training set.

We could try to optimize the threshold for collapsing (see the next set of slides on model tuning).

Does othering help?

hotel_other_wflow <-
  hotel_lm_wflow %>%
  update_recipe(hotel_other_rec)

hotel_other_res <-
  hotel_other_wflow %>%
  fit_resamples(hotel_rs, control = ctrl, metrics = reg_metrics)

collect_metrics(hotel_other_res)
#> # A tibble: 2 × 6
#>   .metric .estimator   mean     n std_err .config             
#>   <chr>   <chr>       <dbl> <int>   <dbl> <chr>               
#> 1 mae     standard   17.4      10 0.205   Preprocessor1_Model1
#> 2 rsq     standard    0.874    10 0.00417 Preprocessor1_Model1

Aabout the same MAE and much faster to complete.

Now let’s look at a more sophisticated tool called effect feature hashing.

Feature Hashing

Between agent and company, simple dummy variables would create 198 new columns (that are mostly zeros).

Another option is to have a binary indicator that combines some levels of these variables.

Feature hashing (for more see FES, SMLTAR, and TMwR):

• uses the character values of the levels
• converts them to integer hash values
• uses the integers to assign them to a specific indicator column.

Feature Hashing

Suppose we want to use 32 indicator variables for agent.

For a agent with value “Max_Kuhn”, a hashing function converts it to an integer (say 210397726).

To assign it to one of the 32 columns, we would use modular arithmetic to assign it to a column:

# For "Max_Kuhn" put a '1' in column: 
210397726 %% 32
#> [1] 30

Hash functions are meant to emulate randomness.

Feature Hashing Pros

• The procedure will automatically work on new values of the predictors.
• It is fast.
• “Signed” hashes add a sign to help avoid aliasing.

Feature Hashing Cons

• There is no real logic behind which factor levels are combined.
• We don’t know how many columns to add (more in the next section).
• Some columns may have all zeros.
• If a indicator column is important to the model, we can’t easily determine why.

The signed hash make it slightly more possible to differentiate between confounded levels

Feature Hashing in recipes

The textrecipes package has a step that can be added to the recipe:

library(textrecipes)

hash_rec <-
  recipe(avg_price_per_room ~ ., data = hotel_tr) %>%
  step_YeoJohnson(lead_time) %>%
  # Defaults to 32 signed indicator columns
  step_dummy_hash(agent) %>%
  step_dummy_hash(company) %>%
  # Regular indicators for the others
  step_dummy(all_nominal_predictors()) %>% 
  step_zv(all_predictors())

hotel_hash_wflow <-
  hotel_lm_wflow %>%
  update_recipe(hash_rec)

Feature Hashing in recipes

hotel_hash_res <-
  hotel_hash_wflow %>%
  fit_resamples(hotel_rs, control = ctrl, metrics = reg_metrics)

collect_metrics(hotel_hash_res)
#> # A tibble: 2 × 6
#>   .metric .estimator   mean     n std_err .config             
#>   <chr>   <chr>       <dbl> <int>   <dbl> <chr>               
#> 1 mae     standard   17.5      10 0.256   Preprocessor1_Model1
#> 2 rsq     standard    0.872    10 0.00395 Preprocessor1_Model1

About the same performance but now we can handle new values.

Debugging a recipe

• Typically, you will want to use a workflow to estimate and apply a recipe.

• If you have an error and need to debug your recipe, the original recipe object (e.g. hash_rec) can be estimated manually with a function called prep(). It is analogous to fit(). See TMwR section 16.4

• Another function (bake()) is analogous to predict(), and gives you the processed data back.

• The tidy() function can be used to get specific results from the recipe.

Example

hash_rec_fit <- prep(hash_rec)

# Get the transformation coefficient
tidy(hash_rec_fit, number = 1)

# Get the processed data
bake(hash_rec_fit, hotel_tr %>% slice(1:3), contains("_agent_"))

More on recipes

• Once fit() is called on a workflow, changing the model does not re-fit the recipe.

• A list of all known steps is at https://www.tidymodels.org/find/recipes/.

• Some steps can be skipped when using predict().

• The order of the steps matters.