Principle:Interpretml Interpret Model Aggregation And Postprocessing

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

Overview

A postprocessing procedure that aggregates bagged model outputs into final term scores with uncertainty estimates and ensures proper term ordering.

Description

After bagged gradient boosting completes, each outer bag produces its own intercept and set of term score tensors. Model Aggregation averages these across bags (weighted by bag weights) to produce the final model, and computes per-term standard deviations for confidence bands. Additional postprocessing includes ANOVA purification (redistributing interaction scores to main effects where possible), ordering terms (main effects first, then interactions), generating human-readable term names, and removing unused bin levels.

Usage

Use this principle after all boosting bags have completed to finalize the model. It should be applied whenever multiple independent model fits need to be combined and the result must be cleaned for presentation.

Theoretical Basis

For B bags with weights w_b:

${\overline{f}}_t(x)=\frac{\sum _{b=1}^B{w}_b\cdot f_t^{(b)}(x)}{\sum _{b=1}^B{w}_b}$

${\sigma }_t(x)=\sqrt{\frac{\sum _{b=1}^B{w}_b\cdot (f_t^{(b)}(x)-{\overline{f}}_t(x){)}^2}{\sum _{b=1}^B{w}_b}}$

ANOVA purification ensures that interaction terms only capture true interaction effects by subtracting main-effect components.

Related Pages

• Implementation:Interpretml_Interpret_Process_Terms