Regional Explanations via Recursive Partitioning

This direction is no longer present in the current build. Its last version stays readable and citable here.

Version 1 of 114 papers26 conceptsfrozen 2026-10-04

Global explanation methods for machine learning—such as Partial Dependence Plots (PDP) or SHAP—often suffer from a fundamental flaw: aggregation bias. By averaging a feature's effect across the entire dataset, these methods can mask complex interactions and heterogeneous behaviors. If a feature has a positive effect on one subgroup but a negative effect on another, the global average might suggest the feature has no effect at all.

To solve this, researchers have developed a framework for regional explanations. The core intuition is to use recursive partitioning (decision trees) not to predict the target variable, but to segment the input space into regions where the model’s behavior is homogeneous and interpretable.

The Problem of Aggregation Bias

When we ask "how does feature $X$ affect the prediction?", we usually look at a global summary. However, as Bernd Bischl, Julia Herbinger, and Giuseppe Casalicchio argue in REPID, these summaries often neutralize conflicting signals.

For example, consider a model predicting house prices where "number of rooms" interacts with "location." In a city center, more rooms might significantly increase price, while in a rural area, the effect might be negligible. A global PDP would show a moderate increase for rooms everywhere, which is technically "correct" on average but practically misleading for any specific house.

Partitioning for Homogeneity

The solution is to find subspaces where the model behaves consistently. This is achieved by applying recursive partitioning to the explanations themselves.

  1. Generate Local Effects: First, calculate local explanations for every data point, such as Individual Conditional Expectation (ICE) curves or Forward Marginal Effects (FMEs).
  2. Split the Space: Use a decision tree algorithm to split the data into subgroups. Instead of minimizing prediction error (like a standard CART tree), the algorithm minimizes the variance of the effect within each leaf.
  3. Regional Plots: Each leaf of the tree represents a "region" of the feature space. Within these regions, the feature effects are now relatively homogeneous, allowing for "Regional Effect Plots" that accurately represent that specific population.

In REPID, the authors optimize splits to minimize the distance between ICE curves. Similarly, the fmeffects package, developed by Holger Löwe and Christian A. Scholbeck, automates this subgroup discovery to reveal where marginal effects differ significantly from the global mean.

Resolving the XAI Disagreement Problem

A secondary benefit of regional partitioning is resolving the "Disagreement Problem" in Explainable AI (XAI). Different methods like SHAP and PDP often provide conflicting feature importance rankings because they handle interactions and correlations differently.

Gabriel Laberge, Yann Pequignot, and colleagues proposed FD-Trees in Tackling the XAI Disagreement Problem with Regional Explanations. They use recursive partitioning to find regions where the complex model is approximately additive. In these regions, interactions are minimized, causing different XAI methods to converge on the same explanation. This idea is further generalized in the GRANITE framework, which identifies regions of "agreement" to provide more trustworthy insights.

Applications and Tools

This approach spans multiple domains:

  • Psychometrics: Achim Zeileis and Carolin Strobl used model-based recursive partitioning in psychotree to identify how latent traits and item difficulties vary across subgroups without pre-specifying them.
  • Medicine: In oncology, Jinshuai Li and colleagues applied sub-regional habitat partitioning to radiological data to capture internal tumor heterogeneity.
  • Software Ecosystems: Tools like ingredients (by Przemyslaw Biecek) and Effector (by Christos Diou and Maximilian Muschalik) provide practitioners with the means to move beyond "one-size-fits-all" explanations.

By shifting the focus from "what the model does on average" to "what the model does in this specific region," recursive partitioning transforms black-box auditing from a search for a single truth into a map of diverse, localized behaviors.

Papers in this version

History & citationVersion 1 · 2026-10-04

No longer present in the current build · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_regional_explanations_via_recursive_part_f5f9840c,
  title = {Regional Explanations via Recursive Partitioning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/regional-explanations-via-recursive-partitioning/rdr_f5f9840c32404be98c75cd4b0f8be943},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_regional_explanations_via_recursive_part_f5f9840c_v1,
  title = {Regional Explanations via Recursive Partitioning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/regional-explanations-via-recursive-partitioning/rdr_f5f9840c32404be98c75cd4b0f8be943/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}