叶节点值作为坐标：梯度提升集成模型的精确对比解释
文章背景与核心概要
梯度提升决策树（GBDT）集成模型在各类表格数据任务中展现出了卓越的预测性能，但其复杂的树结构组合使得模型的可解释性和事后归因（Post-hoc Explanation）变得极具挑战性。现有的许多解释方法依赖于特征层面的加性假设、采样或拟合近似，往往无法精确还原模型的真实决策逻辑。
为了突破这一局限，本文提出了一种新颖的几何解释方法：将梯度提升集成模型中的叶节点值直接视为 \(\mathbb{R}^M\) 空间中的坐标。在这种全新的几何表征下，每个数据样本都被映射为一个线性点，其最终预测得分即为其激活的叶节点坐标之和。这一视角将对比解释（Contrastive Explanation）转化为了精确的算术计算，无需任何拟合、采样或特征可加性假设。在对五个表格数据集的回溯（Recourse）方法评估中，该方法能够以 \(6.2 \times 10^{-15}\) 的精度重构模型决策，在努力程度与真实感之间实现了帕累托非劣效（Pareto-non-dominated）性能，并通过严格过滤可操作的建议，大幅超越了现有的强大基线方法。
Metadata
- arXiv ID: arXiv:2608.19127 [cs.LG]
- Authors: Emanuele Luzio
- Submitted: August 19, 2026
- Subjects: Machine Learning (
cs.LG); Artificial Intelligence (cs.AI); Computers and Society (cs.CY) - ACM Classes: I.2.6; K.4.1
- Citation:
Luzio, E. (2026). Leaf Values as Coordinates: Exact Contrastive Explanation for Gradient-Boosted Ensembles. arXiv:2608.19127.
Abstract
梯度提升集成模型通过对每棵树的一个叶节点值求和来进行预测。如果将这些值解读为坐标而非中间结果，那么每一个样本都变成了 \(\mathbb{R}^M\) 中的一个点，模型在其上呈现线性作用：预测得分即为坐标的总和。
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 \(\mathbb{R}^M\) on which the model acts linearly: the score is the sum of the coordinates.
这种微小的视角转变使得对比解释变得精确。两个样本之间的差异是一个向量，只要它们共享同一个叶节点，该向量在该位置就恒等于零；因此，被拒绝的申请人与被接受的申请人之间的差距由少数几个坐标承载，每个坐标都可以追溯到真实树中的真实分裂。这里不需要进行任何拟合、采样或特征加性假设——加性性质本来就存在于正确的空间之中。
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.
我们基于这种表征构建了一种回溯方法（Recourse Method），并在重复交叉验证下对五个表格数据集进行了评估。其推荐结果能够以 \(6.2 \times 10^{-15}\) 的精度重构模型自身的决策，因此审计人员无需借助模型即可重新核对算术结果。在信用数据集上，它在努力程度与现实感（Effort versus Realism）的权衡上达到了帕累托非劣效（Pareto-non-dominated）。此外，当将推荐限制为主体确实能够改变的因素（例如：不改变年龄、不改变已解决的违约记录）时，它保留了 \(58\%\) 的有效性，而最强的基线方法仅保留了 \(41\%\)。由于标准评估从未询问推荐是否真正可执行，这种差异是标准评估所无法察觉的。
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 \(6.2 \times 10^{-15}\), 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.