---
title: "Leaf Values as Coordinates: Exact Contrastive Explanation for Gradient-Boosted Ensembles"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.19127v1"
permalink: "https://feed.astralor.com/items/itm_ec322f253f4419444cbd2554"
published: "2026-08-19T17:20:33.000Z"
score: 75
channel: "ai"
tags: ["Explainable AI","Recourse","Contrastive Explanation","Gradient-Boosted Ensemble","Tabular Data"]
---

# Leaf Values as Coordinates: Exact Contrastive Explanation for Gradient-Boosted Ensembles

## AI commentary

> The beauty of this paper lies in how a tiny shift in perspective tackles a hard problem, moving contrastive explanation from heuristics to exact verification. For practitioners in credit scoring, fraud detection, and lending, it is not just an academic result but a tool that can directly support audit workflows. I particularly appreciate its emphasis on whether recommendations are actually actionable—a blind spot in current XAI evaluation metrics.

## At a glance

This paper proposes viewing the leaf value of each tree in a gradient-boosted ensemble as a coordinate rather than an intermediate result, mapping every instance to a point in R^M where the model acts linearly. This change of perspective makes contrastive explanation exact: the difference between two instances is nonzero only on coordinates where they do not share a leaf, and each such coordinate is traceable to a real split. The recourse method built on this representation reconstructs the model's own decision to 6.2×10^-15 across five tabular datasets under repeated cross-validation. On credit datasets it is Pareto-non-dominated on effort against realism. When restricted to changes the subject could actually make (excluding age or settled delinquency), it retains 58% validity versus 41% for the strongest baseline, a distinction standard evaluation misses.

## Key points

- Treating leaf values as coordinates gives a linear representation in R^M, making contrastive explanation exact.
- The difference vector between two instances is nonzero only on coordinates where leaves differ, each traceable to an actual split.
- No fitting, sampling, or additive-feature assumption is required; additivity already exists in the coordinate space.
- The recourse method reconstructs model decisions to 6.2×10^-15 on five tabular datasets, retaining 58% validity under actionable-change restrictions.

## Full text

A gradient-boosted ensemble predicts by summing one leaf value per tree. Read those values as coordinates rather than as intermediate results, and every instance becomes a point in R^M on which the model acts linearly: the score is the sum of the coordinates. This small change of view makes contrastive explanation exact. The difference between two instances is a vector that is identically zero wherever they share a leaf, so the gap between a rejected applicant and an accepted one is carried by a handful of coordinates, each traceable to a real split in a real tree. Nothing is fitted, sampled, or assumed additive in features -- the additivity is already there, in the right space. We build a recourse method on this representation and evaluate it on five tabular datasets under repeated cross-validation. Its recommendation reconstructs the model's own decision to 6.2 x 10^-15, so an auditor can re-check the arithmetic without the model. On the credit datasets it is Pareto-non-dominated on effort against realism. And when recommendations are restricted to changes the subject could actually make -- not their age, not a settled delinquency -- it retains 58% of its validity where the strongest baseline retains 41%, a distinction the standard evaluation cannot see because it never asks whether a recommendation can be carried out.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_ec322f253f4419444cbd2554