Principle:Interpretml Interpret EBM Prediction

Metadata
Sources	InterpretML
Domains	Machine_Learning, Interpretability
Last Updated	2026-02-07 12:00 GMT

Overview

A prediction mechanism that computes model outputs by summing per-term score lookups from discretized feature values.

Description

EBM Prediction generates predictions by discretizing each input sample's features into bin indices, looking up the corresponding score in each term's score tensor, and summing all term contributions plus the intercept. The result is a raw score in link-function space. For classification, an inverse link function (logit/softmax) converts scores to probabilities. For regression, the identity or log link is applied. The additive structure enables decomposition of predictions into per-term contributions for local explanations.

Usage

Use this principle whenever you need to generate predictions from a trained EBM. It applies to both batch prediction (predict/predict_proba) and per-sample explanation (explain_local).

Theoretical Basis

For a trained EBM with T terms:

$F(x)={\beta }_0+\sum _{t=1}^T{f}_t(x_{S_t})$

where β₀ is the intercept, f_t is the learned shape function for term t, and S_t is the set of feature indices for that term.

The predicted output:

$\hat{y}=g^{-1}(F(x))$

where g is the link function.

• Classification (binary): g = logit, so $\hat{y}=\sigma (F(x))$
• Regression: g = identity, so $\hat{y}=F(x)$

Related Pages

• Implementation:Interpretml_Interpret_Ebm_Predict_Scores