Implementation:Interpretml Interpret InteractionCore

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

Overview

InteractionCore is a C++ module that manages the creation, initialization, and lifecycle of the interaction detection engine in the EBM framework.

Description

This module implements the core interaction detection functionality through the InteractionCore class. Interaction detection evaluates how strongly pairs (or higher-order tuples) of features interact, which informs which feature interactions should be included in the EBM model.

Key responsibilities include:

• InteractionCore::Create: Factory method that allocates and initializes the interaction core. It processes the shared dataset to extract feature metadata (bin counts, missing/unseen/nominal flags), resolves the objective function, validates targets, checks SIMD and numeric restrictions, unpacks the bag, and initializes the interaction dataset.

• InteractionCore::Free: Thread-safe reference-counted deallocation using std::atomic with proper memory ordering (release on decrement, acquire fence before delete).

• InteractionCore::InitializeInteractionGradientsAndHessians: Computes initial gradients and hessians for all training samples by applying a zero-valued update through the objective function. This reuses the same code path as boosting to generate gradients from initial scores. For classification, it handles both multiclass (with per-class targets stored as unsigned integers) and binary targets. For regression, targets are stored as floats. After gradient computation, it optionally multiplies gradients and hessians by sample weights as an optimization.

• CheckInteractionRestrictions: Validates that bin sizes and index ranges fit within the numeric types used for SIMD processing, checking both UIntBig and UIntSmall variants.

Usage

This module is instantiated when the user calls interaction detection methods. It creates the interaction dataset, initializes gradients, and then serves as the context for all subsequent interaction scoring calls.

Code Reference

Source Location

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

Signature

// Factory method
static ErrorEbm InteractionCore::Create(
    const unsigned char* const pDataSetShared,
    const size_t cSamples,
    const size_t cFeatures,
    const size_t cWeights,
    const BagEbm* const aBag,
    const CreateInteractionFlags flags,
    const AccelerationFlags acceleration,
    const char* const sObjective,
    const double* const experimentalParams,
    InteractionCore** const ppInteractionCoreOut);

// Reference-counted destructor
static void InteractionCore::Free(
    InteractionCore* const pInteractionCore);

// Gradient/Hessian initialization
ErrorEbm InteractionCore::InitializeInteractionGradientsAndHessians(
    const unsigned char* const pDataSetShared,
    const size_t cWeights,
    const double* const aIntercept,
    const BagEbm* const aBag,
    const double* const aInitScores);

I/O Contract

Inputs

Name	Type	Required	Description
pDataSetShared	const unsigned char*	Yes	Shared dataset binary blob
cSamples	size_t	Yes	Total number of samples
cFeatures	size_t	Yes	Number of features
cWeights	size_t	Yes	Number of weight columns (0 or 1)
aBag	const BagEbm*	No	Bag replication array for train/validation split
flags	CreateInteractionFlags	Yes	Configuration flags (UseApprox, DifferentialPrivacy, BinaryAsMulticlass)
sObjective	const char*	Yes	Objective function name string
aIntercept	const double*	No	Initial intercept values per score
aInitScores	const double*	No	Initial score values per sample

Outputs

Name	Type	Description
return value	ErrorEbm	Error code (Error_None on success)
ppInteractionCoreOut	InteractionCore**	Pointer to the created InteractionCore instance

Usage Examples

Pipeline Context

# This C++ module is called internally via the native bindings
# during interaction detection to score feature interactions
from interpret.glassbox import ExplainableBoostingClassifier
ebm = ExplainableBoostingClassifier(interactions=10)
ebm.fit(X, y)  # Internally creates InteractionCore for interaction scoring

Related Pages

• Environment:Interpretml_Interpret_Native_Libebm_Environment