Interpretability-by-Design with Accurate Locally Additive Models and Conditional Feature Effects

arXiv:2602.16503 · cs.LG, cs.AI · Submitted 2026-02-18 · Read on arXiv

cs.LG, cs.AI

Submitted: 2026-02-18

Updated: 2026-11-06

Comments: Accepted at NeurIPS 2026. 41 pages, 10 figures. Code: https://github.com/givasile/calm-paper

Code: https://github.com/givasile/CALM

License: http://creativecommons.org/licenses/by/4.0/

The gist: Generalized additive models (GAMs) offer high interpretability through independent univariate feature effects but underfit when interactions are present in data.

Terminology

Abstract

Generalized additive models (GAMs) offer high interpretability through independent univariate feature effects but underfit when interactions are present in data. GA squared Ms 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 GA squared Ms 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 GA squared Ms, while preserving the univariate auditability that GA squared Ms forfeit.

Sources

Related papers