Implementation:Interpretml Interpret PartitionMultiDimensionalFull

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

Overview

Computes the full partition gain for multi-dimensional interaction terms by iterating over all tensor bins and calculating the improvement over the baseline.

Description

The PartitionMultiDimensionalFull.cpp file implements the PartitionMultiDimensionalFullInternal and PartitionMultiDimensionalFullTarget template classes. The core Func method iterates over all bins in a multi-dimensional histogram tensor, accumulating total gradient/hessian/weight/count statistics while simultaneously computing the per-bin partial gain using CalcPartialGain (with L1/L2 regularization and delta step capping). The final gain is computed by subtracting the total-bin gain from the sum of per-bin gains, measuring how much better individual bins predict compared to a single overall prediction. The function supports both Newton (hessian-based) and gradient boosting modes, selected via the CalcInteractionFlags_DisableNewton flag.

Usage

Called during interaction detection to compute the FAST interaction strength for a candidate feature pair. The gain value returned indicates how much predictive improvement can be obtained by including this interaction term in the model.

Code Reference

Source Location

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

Signature

template<bool bHessian, size_t cCompilerScores>
class PartitionMultiDimensionalFullInternal final {
public:
   INLINE_RELEASE_UNTEMPLATED static double Func(
       InteractionCore* const pInteractionCore,
       const size_t cTensorBins,
       const CalcInteractionFlags flags,
       const FloatCalc regAlpha,
       const FloatCalc regLambda,
       const FloatCalc deltaStepMax,
       BinBase* const aAuxiliaryBinsBase,
       BinBase* const aBinsBase);
};

template<bool bHessian, size_t cPossibleScores>
class PartitionMultiDimensionalFullTarget final {
public:
   INLINE_RELEASE_UNTEMPLATED static double Func(
       InteractionCore* const pInteractionCore,
       const size_t cTensorBins,
       const CalcInteractionFlags flags,
       const FloatCalc regAlpha,
       const FloatCalc regLambda,
       const FloatCalc deltaStepMax,
       BinBase* const aAuxiliaryBinsBase,
       BinBase* const aBinsBase);
};

I/O Contract

Input	Description
pInteractionCore	Core interaction state with score count
cTensorBins	Total number of bins in the multi-dimensional tensor
flags	CalcInteractionFlags (e.g., DisableNewton)
regAlpha	L1 regularization parameter
regLambda	L2 regularization parameter
deltaStepMax	Maximum step size cap
aBinsBase	Populated histogram bins

Output	Description
double return	Computed gain (interaction strength)

Usage Examples

# Called internally via native bindings
from interpret.glassbox import ExplainableBoostingClassifier
ebm = ExplainableBoostingClassifier(interactions=10)
ebm.fit(X, y)  # PartitionMultiDimensionalFull computes interaction gains

Related Pages

• Environment:Interpretml_Interpret_Native_Libebm_Environment