OptimalBinningWoE (version 1.0.8)

step_obwoe: Optimal Binning and WoE Transformation Step

Description

step_obwoe() creates a specification of a recipe step that discretizes predictor variables using one of 28 state-of-the-art optimal binning algorithms and transforms them into Weight of Evidence (WoE) values. This step fully integrates the OptimalBinningWoE package into the tidymodels framework, supporting supervised discretization for both binary and multinomial classification targets with extensive hyperparameter tuning capabilities.

Usage

step_obwoe(
  recipe,
  ...,
  role = "predictor",
  trained = FALSE,
  outcome = NULL,
  algorithm = "auto",
  min_bins = 2L,
  max_bins = 10L,
  bin_cutoff = 0.05,
  output = c("woe", "bin", "both"),
  suffix_woe = "_woe",
  suffix_bin = "_bin",
  na_woe = 0,
  control = list(),
  binning_results = NULL,
  skip = FALSE,
  id = recipes::rand_id("obwoe")
)

Value

An updated recipe object with the new step appended.

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 variables for this step. See selections for available selectors. Common choices include all_predictors(), all_numeric_predictors(), or all_nominal_predictors(). Ensure the selected variables are compatible with the chosen algorithm (e.g., do not apply "mdlp" to categorical data).

role

For variables created by this step, what role should they have?

Default is "predictor".

trained

A logical indicating whether the step has been trained (fitted). This should not be set manually.

outcome

A character string specifying the name of the binary or multinomial response variable. This argument is required as all binning algorithms are supervised. The outcome must exist in the training data provided to prep(). The outcome should be encoded as a factor (standard for tidymodels classification) or as integers 0/1 for binary, 0/1/2/... for multinomial.

algorithm

Character string specifying the binning algorithm to use. Use "auto" (default) for automatic selection based on target type: "jedi" for binary targets, "jedi_mwoe" for multinomial.

Available algorithms are organized by supported feature types:

Universal (numerical and categorical): "auto", "jedi", "jedi_mwoe", "cm", "dp", "dmiv", "fetb", "mob", "sketch", "udt"

Numerical only: "bb", "ewb", "fast_mdlp", "ir", "kmb", "ldb", "lpdb", "mblp", "mdlp", "mrblp", "oslp", "ubsd"

Categorical only: "gmb", "ivb", "mba", "milp", "sab", "sblp", "swb"

This parameter is tunable with tune().

min_bins

Integer specifying the minimum number of bins to create. Must be at least 2. Default is 2. This parameter is tunable with tune().

max_bins

Integer specifying the maximum number of bins to create. Must be greater than or equal to min_bins. Default is 10. This parameter is tunable with tune().

bin_cutoff

Numeric value between 0 and 1 (exclusive) specifying the minimum proportion of total observations that each bin must contain. Bins with fewer observations are merged with adjacent bins. This serves as a regularization mechanism to prevent overfitting and ensure statistical stability of WoE estimates. Default is 0.05 (5%). This parameter is tunable with tune().

output

Character string specifying the transformation output format:

"woe"

Replaces the original variable with WoE values (default). This is the standard choice for logistic regression scorecards.

"bin"

Replaces the original variable with bin labels (character). Useful for tree-based models or exploratory analysis.

"both"

Keeps the original column and adds two new columns with suffixes _woe and _bin. Useful for model comparison or audit trails.

suffix_woe

Character string suffix appended to create WoE column names when output = "both". Default is "_woe".

suffix_bin

Character string suffix appended to create bin column names when output = "both". Default is "_bin".

na_woe

Numeric value to assign to observations that cannot be mapped to a bin during bake(). This includes missing values (NA) and unseen categories not present in the training data. Default is 0, which represents neutral evidence (neither good nor bad).

control

A named list of additional control parameters passed to control.obwoe. These provide fine-grained control over algorithm behavior such as convergence thresholds and maximum pre-bins. Parameters specified directly in step_obwoe() (e.g., bin_cutoff) take precedence over values in this list.

binning_results

Internal storage for fitted binning models after prep(). Do not set manually.

skip

Logical. Should this step be skipped when bake() is called on new data? Default is FALSE. Setting to TRUE is rarely needed for WoE transformations but may be useful in specialized workflows.

id

A unique character string to identify this step. If not provided, a random identifier is generated.

Details

Weight of Evidence Transformation

Weight of Evidence (WoE) is a supervised encoding technique that transforms categorical and continuous variables into a scale that measures the predictive strength of each value or bin relative to the target variable. For a bin \(i\), the WoE is defined as:

$$WoE_i = \ln\left(\frac{\text{Distribution of Events}_i}{\text{Distribution of Non-Events}_i}\right)$$

Positive WoE values indicate the bin has a higher proportion of events (e.g., defaults) than the overall population, while negative values indicate lower risk.

Algorithm Selection Strategy

The algorithm parameter provides access to 28 binning algorithms:

• Use algorithm = "auto" (default) for automatic selection: "jedi" for binary targets, "jedi_mwoe" for multinomial.

• Use algorithm = "mob" (Monotonic Optimal Binning) when monotonic WoE trends are required for regulatory compliance (Basel/IFRS 9).

• Use algorithm = "mdlp" for entropy-based discretization of numerical variables (requires all_numeric_predictors()).

• Use algorithm = "dp" (Dynamic Programming) for exact optimal solutions when computational cost is acceptable.

If an incompatible algorithm is applied to a variable (e.g., "mdlp" on a factor), the step will issue a warning during prep() and skip that variable, leaving it untransformed.

Handling New Data

During bake(), observations are mapped to bins learned during prep():

• Numerical variables: Values are assigned to bins based on the learned cutpoints using interval notation.

• Categorical variables: Categories are matched to their corresponding bins. Categories not seen during training receive the na_woe value.

• Missing values: Always receive the na_woe value.

Tuning with tune

This step is fully compatible with the tune package. The following parameters support tune():

• algorithm: See obwoe_algorithm.

• min_bins: See obwoe_min_bins.

• max_bins: See obwoe_max_bins.

• bin_cutoff: See obwoe_bin_cutoff.

Case Weights

This step does not currently support case weights. All observations are treated with equal weight during binning optimization.

References

Siddiqi, N. (2006). Credit Risk Scorecards: Developing and Implementing Intelligent Credit Scoring. John Wiley & Sons. tools:::Rd_expr_doi("10.1002/9781119201731")

Thomas, L. C., Edelman, D. B., & Crook, J. N. (2002). Credit Scoring and Its Applications. SIAM Monographs on Mathematical Modeling and Computation. tools:::Rd_expr_doi("10.1137/1.9780898718317")

Navas-Palencia, G. (2020). Optimal Binning: Mathematical Programming Formulation and Solution Approach. Expert Systems with Applications, 158, 113508. tools:::Rd_expr_doi("10.1016/j.eswa.2020.113508")

See Also

obwoe for the underlying binning engine, control.obwoe for advanced control parameters, obwoe_algorithm, obwoe_min_bins, obwoe_max_bins, obwoe_bin_cutoff for tuning parameter definitions, recipe, prep, bake for the tidymodels recipe framework.

Examples

# \donttest{
library(recipes)

# Simulated credit data
set.seed(123)
credit_data <- data.frame(
  age = rnorm(500, 45, 12),
  income = exp(rnorm(500, 10, 0.6)),
  employment = sample(c("Employed", "Self-Employed", "Unemployed"),
    500,
    replace = TRUE, prob = c(0.7, 0.2, 0.1)
  ),
  education = factor(c("HighSchool", "Bachelor", "Master", "PhD")[
    sample(1:4, 500, replace = TRUE, prob = c(0.3, 0.4, 0.2, 0.1))
  ]),
  default = factor(rbinom(500, 1, 0.15),
    levels = c(0, 1),
    labels = c("No", "Yes")
  )
)

# Example 1: Basic usage with automatic algorithm selection
rec_basic <- recipe(default ~ ., data = credit_data) %>%
  step_obwoe(all_predictors(), outcome = "default")

rec_prepped <- prep(rec_basic)
baked_data <- bake(rec_prepped, new_data = NULL)
head(baked_data)

# View binning details
tidy(rec_prepped, number = 1)

# Example 2: Numerical-only algorithm on numeric predictors
rec_mdlp <- recipe(default ~ age + income, data = credit_data) %>%
  step_obwoe(all_numeric_predictors(),
    outcome = "default",
    algorithm = "mdlp",
    min_bins = 3,
    max_bins = 6
  )

# Example 3: Output both bins and WoE
rec_both <- recipe(default ~ age, data = credit_data) %>%
  step_obwoe(age,
    outcome = "default",
    output = "both"
  )

baked_both <- bake(prep(rec_both), new_data = NULL)
names(baked_both)
# Contains: default, age, age_woe, age_bin

# Example 4: Custom control parameters
rec_custom <- recipe(default ~ ., data = credit_data) %>%
  step_obwoe(all_predictors(),
    outcome = "default",
    algorithm = "mob",
    bin_cutoff = 0.03,
    control = list(
      max_n_prebins = 30,
      convergence_threshold = 1e-8
    )
  )

# Example 5: Tuning specification (for use with tune package)
# rec_tune <- recipe(default ~ ., data = credit_data) %>%
#   step_obwoe(all_predictors(),
#              outcome = "default",
#              algorithm = tune(),
#              min_bins = tune(),
#              max_bins = tune())
# }