SAILS: Surrogate-based Analysis of Interactions via Local Effect Smooths

2026-06-08

AI Summary

Existing explainable AI (XAI) methods struggle to characterize the precise functional forms of feature interactions, typically supporting only limited types of detection or visualization. This work proposes the SAILS framework, which, for the first time in a model-agnostic setting, fits interpretable generalized additive model (GAM) surrogates to local effects and decomposes pairwise feature interactions at the derivative level. The method systematically categorizes these interactions into three types—linear, multiplicatively separable, and non-multiplicatively separable—and provides tailored visualizations for each. By integrating local smoothing, derivative-based interaction decomposition, statistical significance testing, and interaction-type classification, SAILS effectively uncovers interaction mechanisms in both synthetic and real-world tasks, thereby addressing a critical gap in XAI regarding the modeling of interaction functional forms. Nevertheless, limitations remain in scenarios involving highly correlated features or higher-order interactions.

Abstract

Feature interactions drive much of the predictive power of machine learning models, yet existing explanation methods only detect and quantify interactions without revealing their functional form, or visualize only restricted interaction types. We propose Surrogate-based Analysis of Interactions via Local effect Smooths (SAILS), a model-agnostic framework that analyzes pairwise interactions through interpretable generalized additive model (GAM) surrogates fitted to the local effects of a black-box model. For each interval of a feature of interest, the surrogate smooth terms isolate the interaction components on derivative level, enabling (i) interaction detection through a heuristic derived from significance tests on smooth terms, (ii) interaction form categorization into linear, product-separable, and non-product-separable types, and (iii) tailored, interpretable visualizations for each interaction type. We empirically validate the framework through controlled simulations and a real-world task, demonstrating its effectiveness for pairwise interactions, with limitations under strong feature correlations and higher-order interactions. SAILS fills a notable gap in the XAI toolbox, going beyond detection of interactions alone to characterizing their functional form.

Problem

• feature interactions
• interpretability
• explainable AI
• functional form
• model-agnostic

Innovation

• feature interactions
• model-agnostic
• generalized additive models
• local effect smooths
• interpretable machine learning