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

Researchers introduce SAILS, a model-agnostic framework that goes beyond detecting feature interactions in machine learning models to reveal their functional forms and characteristics. Using surrogate generalized additive models, SAILS categorizes interactions as linear, product-separable, or non-product-separable and provides tailored visualizations, advancing the field of explainable AI.

SAILS addresses a critical gap in explainable AI (XAI) by moving beyond interaction detection to functional characterization. Machine learning models derive much predictive power from feature interactions, yet existing explanation methods either identify that interactions exist without revealing how they work, or visualize only restricted interaction types. This research bridges that interpretability divide through a sophisticated surrogate-based approach that fits generalized additive models (GAMs) to capture the local effects of black-box models.

The framework's innovation lies in its ability to isolate interaction components at the derivative level for specific feature intervals, enabling both statistical detection and categorical classification. By categorizing interactions into linear, product-separable, and non-product-separable types, SAILS provides practitioners with actionable insights into model behavior. The research validates this approach through controlled simulations and real-world applications, demonstrating effectiveness for pairwise interactions while acknowledging limitations with highly correlated features and higher-order interactions.

For the AI and machine learning communities, SAILS represents progress toward more transparent, interpretable models—a growing necessity as ML systems face increased regulatory scrutiny and deployment in high-stakes domains. Financial institutions, healthcare providers, and other regulated sectors depend on model explainability for compliance and trust-building. The framework's model-agnostic design means it applies across diverse architectures and use cases.

Future development should focus on extending SAILS to handle higher-order interactions and feature correlations more robustly. As explainability becomes a competitive differentiator and regulatory requirement, tools like SAILS will likely see broader adoption in both research and production environments, particularly in domains where understanding model decisions is non-negotiable.

• →SAILS enables characterization of feature interaction functional forms, not just detection
• →The framework categorizes interactions into linear, product-separable, and non-product-separable types with tailored visualizations
• →Works model-agnostically through surrogate GAM fitting to black-box model local effects
• →Validated on controlled simulations and real-world tasks but has limitations with correlated features and higher-order interactions
• →Addresses a significant gap in XAI tooling relevant to regulated industries requiring model transparency