Implementation:Interpretml Interpret PartitionRandomBoosting

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

Overview

PartitionRandomBoosting is a C++ module that performs random split selection for EBM boosting, primarily used for training differentially private EBM models.

Description

This module implements a random partitioning strategy for boosting interaction terms, where split points are chosen randomly rather than optimally. This is the primary splitting algorithm used when differential privacy is enabled, since optimal split selection would leak information about the training data.

The algorithm works as follows:

1. Collapse degenerate dimensions: Dimensions with only 1 bin are collapsed, and the remaining real dimensions are processed.
2. Determine splits per dimension: For each dimension, computes the number of random splits based on aLeavesMax, capping at the number of available bins.
3. Allocate and generate random splits: Uses the RandomDeterministic RNG to generate random split positions for each dimension, then sorts them to ensure ascending order.
4. Construct the collapsed tensor: Iterates over all cells in the collapsed tensor, mapping each cell to its corresponding original bin and accumulating gradient/hessian statistics.
5. Compute update scores: For each cell in the collapsed tensor, computes the boosting update using regularization (alpha, lambda), Newton or gradient-only updates, and step size constraints. Handles monotone constraints for single-dimension terms.
6. Build the output tensor: Maps the collapsed tensor splits back to the original tensor dimensions and sets the scores in the output Tensor.

The module handles both hessian-aware (Newton) updates and gradient-only updates based on the TermBoostFlags.

Usage

This module is called during boosting when random splitting is selected, which is primarily for differential privacy scenarios. It provides the split selection mechanism that avoids data-dependent split point optimization.

Code Reference

Source Location

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

Signature

template<bool bHessian, size_t cCompilerScores>
class PartitionRandomBoostingInternal final {
public:
    static ErrorEbm Func(
        RandomDeterministic* const pRng,
        BoosterShell* const pBoosterShell,
        const Term* const pTerm,
        const TermBoostFlags flags,
        const FloatCalc regAlpha,
        const FloatCalc regLambda,
        const FloatCalc deltaStepMax,
        const IntEbm* const aLeavesMax,
        const MonotoneDirection monotoneDirection,
        double* const pTotalGain);
};

I/O Contract

Inputs

Name	Type	Required	Description
pRng	RandomDeterministic*	Yes	Deterministic random number generator
pBoosterShell	BoosterShell*	Yes	Boosting context containing bins and model state
pTerm	const Term*	Yes	The interaction term being boosted
flags	TermBoostFlags	Yes	Boosting flags (DisableNewtonGain, DisableNewtonUpdate)
regAlpha	FloatCalc	Yes	L1 regularization parameter
regLambda	FloatCalc	Yes	L2 regularization parameter
deltaStepMax	FloatCalc	Yes	Maximum update step size
aLeavesMax	const IntEbm*	Yes	Maximum leaves per dimension
monotoneDirection	MonotoneDirection	Yes	Monotone constraint direction (NONE for interactions)

Outputs

Name	Type	Description
return value	ErrorEbm	Error code (Error_None on success)
pTotalGain	double*	Total gain from the random split
pBoosterShell (internal)	BoosterShell*	Updated inner term update tensor

Usage Examples

Pipeline Context

# This C++ module is called internally via the native bindings
# primarily when training with differential privacy
from interpret.glassbox import DPExplainableBoostingClassifier
ebm = DPExplainableBoostingClassifier(epsilon=1.0)
ebm.fit(X, y)  # Internally calls PartitionRandomBoosting for DP-safe splits

Related Pages

• Environment:Interpretml_Interpret_Native_Libebm_Environment