[论文解读] Interpretability-by-Design with Accurate Locally Additive Models and Conditional Feature Effects

Vasilis Gkolemis, Loukas Kavouras+6 more|arXiv (Cornell University)|Feb 18, 2026

Explainable Artificial Intelligence (XAI)被引用 0

一句话总结

CALMs 在 GAM 可解释性和 GA2M 准确性之间取得平衡，通过学习区域特异的单变量特征效应来捕捉交互。

ABSTRACT

Generalized additive models (GAMs) offer interpretability through independent univariate feature effects but underfit when interactions are present in data. GA$^2$Ms add selected pairwise interactions which improves accuracy, but sacrifices interpretability and limits model auditing. We propose \emph{Conditionally Additive Local Models} (CALMs), a new model class, that balances the interpretability of GAMs with the accuracy of GA$^2$Ms. CALMs allow multiple univariate shape functions per feature, each active in different regions of the input space. These regions are defined independently for each feature as simple logical conditions (thresholds) on the features it interacts with. As a result, effects remain locally additive while varying across subregions to capture interactions. We further propose a principled distillation-based training pipeline that identifies homogeneous regions with limited interactions and fits interpretable shape functions via region-aware backfitting. Experiments on diverse classification and regression tasks show that CALMs consistently outperform GAMs and achieve accuracy comparable with GA$^2$Ms. Overall, CALMs offer a compelling trade-off between predictive accuracy and interpretability.

研究动机与目标

• 为高风险决策制定可解释设计的模型提供动机。
• 解决存在交互时 GAM 的拟合不足以及 GA2M 的可解释性下降问题。
• 提出条件可加性局部模型（CALMs），以融合可解释性与准确性。
• 开发基于蒸馏的训练流程，识别接近加性区域并学习区域特定的形状函数。

提出的方法

• 将 CALM 定义为带区域条件化激活的单变量形状函数之和：g(E[Y|X=x]) = beta0 + sum_i f_i^{(r_i(x_-i))}(x_i)。
• 使用二叉树将输入空间按特征划分为区域，使得 x_i 在该区域内近似线性相加。
• 使用异质性度量来引导区域分裂，以在区域内最小化交互。
• 在每个区域内通过区域感知回拟合拟合区域特异的单变量形状函数 f_i^{(r)}。
• 三步蒸馏式训练： (1) 训练一个黑盒参考模型， (2) 学习特征特定的分区， (3) 估计区域特异的形状函数。

Figure 1 : CALMs use conditional feature effects : every feature effect is expressed by a collection of 1D functions, each associated with a different region of the input space. In the example, the effect of $x_{1}$ conditions on $x_{3}$ , the effect of $x_{2}$ on $x_{1}$ , while $x_{d}$ does not in

实验结果

研究问题

• RQ1 CALMs 是否能在保持 GAM 类可解释性的同时达到 GA2M 级别的准确性？
• RQ2 如何可靠识别接近加性的区域以捕捉交互而不过度增加复杂性？
• RQ3 相较于 GAM 和 GA2M，CALM 提供怎样的可解释性保证？
• RQ4 所提出的基于蒸馏的训练流程在不同数据集上是否高效且可扩展？

主要发现

• CALMs 在多样任务中始终优于 GAM 的预测准确性。
• CALMs 在许多数据集上达到与 GA2M 相当的准确性。
• CALMs 通过区域条件化的单变量效应提供可解释性，可用于审计。
• 训练流程高效识别交互特征和区域阈值，调参最少。
• 在 25 个公开表格数据集（分类和回归）上的实验支持所提出的准确性与可解释性之间的权衡。

Figure 2 : CALM plot for $x_{1}$ . Each curve shows the contribution of $x_{1}$ to $y$ (P1) in a different region of the input space ( $x_{3}\lessgtr 0$ ). Vertical lines indicate interaction-induced discontinuities, e.g., at $x_{1}\approx-0.4$ there exist a positive hidden jump of $[0,0.37]$ due to