← 研踪 PaperTrailarXiv:2602.16503 · v3
Interpretability-by-Design with Accurate Locally Additive Models and Conditional Feature Effects
首次提交 2026-02-18 22:45
版本更新 2026-11-06 20:55
雷达观测 2026-11-09 11:19
摘要
Generalized additive models (GAMs) offer high interpretability through independent univariate feature effects but underfit when interactions are present in data. GAMs add selected pairwise interactions, which improves accuracy, but sacrifices interpretability. We propose Conditionally Additive Local Models (CALMs), a new model class that narrows this trade-off: it stays close to GAMs in interpretability and close to GAMs in accuracy. 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 never underperform GAMs, often beat them and achieve accuracy broadly comparable to GAMs, while preserving the univariate auditability that GAMs forfeit.
作者备注:Accepted at NeurIPS 2026. 41 pages, 10 figures. Code: https://github.com/givasile/calm-paper
观测到的版本
相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。
v3 · 2026-11-06 20:55 · 最新观测版本
Interpretability-by-Design with Accurate Locally Additive Models and Conditional Feature Effects
Generalized additive models (GAMs) offer high interpretability through independent univariate feature effects but underfit when interactions are present in data. GAMs add selected pairwise interactions, which improves accuracy, but sacrifices interpretability. We propose Conditionally Additive Local Models (CALMs), a new model class that narrows this trade-off: it stays close to GAMs in interpretability and close to GAMs in accuracy. 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 never underperform GAMs, often beat them and achieve accuracy broadly comparable to GAMs, while preserving the univariate auditability that GAMs forfeit.
查看此版本解读 →查看此版本原文 ↗