Implementation:Interpretml Interpret CutWinsorized

Knowledge Sources	Interpretml Interpret
Domains	Machine_Learning, EBM_Core
Last Updated	2026-02-07 12:00 GMT

Overview

CutWinsorized is a C++ module that generates winsorized cut points for feature discretization by trimming outliers from the tails of the data distribution before placing evenly spaced cuts.

Description

The CutWinsorized function implements a winsorized binning strategy where extreme values are trimmed and cut points are placed within the resulting range. The algorithm works by:

1. Copying and sorting the input feature values, removing missing values and replacing infinities with the maximum/minimum representable float values.
2. Determining outer boundary values by moving inward from the sorted extremes by a fraction proportional to the number of bins ((cSamples - 1) / cBins).
3. Finding transition points near the center of the data when only one cut is requested.
4. For multiple cuts, finding the inner transition values and placing evenly spaced cuts between them.
5. Using the ArithmeticMean helper to compute midpoints between transition values.
6. Using FloatTickIncrement to ensure the upper boundary slightly exceeds the last observed value for proper lower-bound-inclusive bin assignment.

The algorithm handles edge cases including single-valued data, data with few distinct values, and numerical overflow in step size computation.

Usage

This module is called during the feature discretization phase when the winsorized binning strategy is selected. It provides a more robust alternative to uniform binning by reducing the influence of extreme outliers on bin boundary placement.

Code Reference

Source Location

• Repository: Interpretml_Interpret
• File: shared/libebm/CutWinsorized.cpp

Signature

EBM_API_BODY ErrorEbm EBM_CALLING_CONVENTION CutWinsorized(
    IntEbm countSamples,
    const double* featureVals,
    IntEbm* countCutsInOut,
    double* cutsLowerBoundInclusiveOut);

I/O Contract

Inputs

Name	Type	Required	Description
countSamples	IntEbm	Yes	Number of feature value samples
featureVals	const double*	Yes	Array of feature values (may contain NaN for missing)
countCutsInOut	IntEbm*	Yes	Pointer to the desired number of cuts (updated on output)
cutsLowerBoundInclusiveOut	double*	Yes	Output buffer for cut points

Outputs

Name	Type	Description
return value	ErrorEbm	Error code (Error_None on success)
countCutsInOut	IntEbm*	Updated with the actual number of cuts placed
cutsLowerBoundInclusiveOut	double*	Array of lower-bound-inclusive cut points

Usage Examples

Pipeline Context

# This C++ module is called internally via the native bindings
# during preprocessing when winsorized binning is selected
from interpret.glassbox import ExplainableBoostingClassifier
ebm = ExplainableBoostingClassifier()
ebm.fit(X, y)  # Internally calls CutWinsorized during discretization

Related Pages

• Environment:Interpretml_Interpret_Native_Libebm_Environment