step_woe: Weight of evidence transformation

In embed: Extra Recipes for Encoding Predictors

Weight of evidence transformation

Description

step_woe() creates a specification of a recipe step that will transform nominal data into its numerical transformation based on weights of evidence against a binary outcome.

Usage

step_woe(
  recipe,
  ...,
  role = "predictor",
  outcome,
  trained = FALSE,
  dictionary = NULL,
  Laplace = 1e-06,
  prefix = "woe",
  keep_original_cols = FALSE,
  skip = FALSE,
  id = rand_id("woe")
)

Arguments

recipe	A recipe object. The step will be added to the sequence of operations for this recipe.
...	One or more selector functions to choose which variables will be used to compute the components. See selections() for more details. For the tidy method, these are not currently used.
role	For model terms created by this step, what analysis role should they be assigned?. By default, the function assumes that the new woe components columns created by the original variables will be used as predictors in a model.
outcome	The bare name of the binary outcome encased in vars().
trained	A logical to indicate if the quantities for preprocessing have been estimated.
dictionary	A tbl. A map of levels and woe values. It must have the same layout than the output returned from dictionary(). If NULL the function will build a dictionary with those variables passed to .... See dictionary() for details.
Laplace	The Laplace smoothing parameter. A value usually applied to avoid -Inf/Inf from predictor category with only one outcome class. Set to 0 to allow Inf/-Inf. The default is 1e-6. Also known as 'pseudocount' parameter of the Laplace smoothing technique.
prefix	A character string that will be the prefix to the resulting new variables. See notes below.
keep_original_cols	A logical to keep the original variables in the output. Defaults to FALSE.
skip	A logical. Should the step be skipped when the recipe is baked by recipes::bake()? While all operations are baked when recipes::prep() is run, some operations may not be able to be conducted on new data (e.g. processing the outcome variable(s)). Care should be taken when using skip = TRUE as it may affect the computations for subsequent operations
id	A character string that is unique to this step to identify it.

Details

WoE is a transformation of a group of variables that produces a new set of features. The formula is

woe_c = log((P(X = c|Y = 1))/(P(X = c|Y = 0)))

where c goes from 1 to C levels of a given nominal predictor variable X.

These components are designed to transform nominal variables into numerical ones with the property that the order and magnitude reflects the association with a binary outcome. To apply it on numerical predictors, it is advisable to discretize the variables prior to running WoE. Here, each variable will be binarized to have woe associated later. This can achieved by using step_discretize().

The argument Laplace is an small quantity added to the proportions of 1's and 0's with the goal to avoid log(p/0) or log(0/p) results. The numerical woe versions will have names that begin with woe_ followed by the respective original name of the variables. See Good (1985).

One can pass a custom dictionary tibble to step_woe(). It must have the same structure of the output from dictionary() (see examples). If not provided it will be created automatically. The role of this tibble is to store the map between the levels of nominal predictor to its woe values. You may want to tweak this object with the goal to fix the orders between the levels of one given predictor. One easy way to do this is by tweaking an output returned from dictionary().

Value

An updated version of recipe with the new step added to the sequence of existing steps (if any). For the tidy method, a tibble with the woe dictionary used to map categories with woe values.

Tidying

When you tidy() this step, a tibble with columns terms (the selectors or variables selected), value, n_tot, n_bad, n_good, p_bad, p_good, woe and outcome is returned.. See dictionary() for more information.

When you tidy() this step, a tibble is retruned with columns terms value, n_tot, n_bad, n_good, p_bad, p_good, woe and outcome and id:

terms

character, the selectors or variables selected

value

character, level of the outcome

n_tot

integer, total number

n_bad

integer, number of bad examples

n_good

integer, number of good examples

p_bad

numeric, p of bad examples

p_good

numeric, p of good examples

woe

numeric, weight of evidence

outcome

character, name of outcome variable

id

character, id of this step

Tuning Parameters

This step has 1 tuning parameters:

• Laplace: Laplace Correction (type: double, default: 1e-06)

Case weights

The underlying operation does not allow for case weights.

References

Kullback, S. (1959). Information Theory and Statistics. Wiley, New York.

Hastie, T., Tibshirani, R. and Friedman, J. (1986). Elements of Statistical Learning, Second Edition, Springer, 2009.

Good, I. J. (1985), "Weight of evidence: A brief survey", Bayesian Statistics, 2, pp.249-270.

Examples

library(modeldata)
data("credit_data")

set.seed(111)
in_training <- sample(1:nrow(credit_data), 2000)

credit_tr <- credit_data[in_training, ]
credit_te <- credit_data[-in_training, ]

rec <- recipe(Status ~ ., data = credit_tr) %>%
  step_woe(Job, Home, outcome = vars(Status))

woe_models <- prep(rec, training = credit_tr)

# the encoding:
bake(woe_models, new_data = credit_te %>% slice(1:5), starts_with("woe"))
# the original data
credit_te %>%
  slice(1:5) %>%
  dplyr::select(Job, Home)
# the details:
tidy(woe_models, number = 1)

# Example of custom dictionary + tweaking
# custom dictionary
woe_dict_custom <- credit_tr %>% dictionary(Job, Home, outcome = "Status")
woe_dict_custom[4, "woe"] <- 1.23 # tweak

# passing custom dict to step_woe()
rec_custom <- recipe(Status ~ ., data = credit_tr) %>%
  step_woe(
    Job, Home,
    outcome = vars(Status), dictionary = woe_dict_custom
  ) %>%
  prep()

rec_custom_baked <- bake(rec_custom, new_data = credit_te)
rec_custom_baked %>%
  dplyr::filter(woe_Job == 1.23) %>%
  head()

embed documentation built on May 29, 2024, 9:59 a.m.