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

SAILS addresses a critical gap in AI explainability: While existing methods can detect that two features interact in a model's predictions, they cannot reveal how they interact. This framework goes further by categorizing interactions into specific types (linear, product-separable, and non-product-separable), providing deeper insight into model behavior.

The method works across different types of AI models: As a "model-agnostic" framework, SAILS can be applied to any black-box machine learning model regardless of its architecture, making it broadly applicable for auditing and understanding diverse AI systems used in practice.

Interpretable visualizations make interactions understandable to non-experts: By creating tailored, visual representations for different interaction types, SAILS translates complex mathematical relationships into formats that stakeholders without technical backgrounds can understand and use to evaluate model fairness and reliability.

The framework has known limitations that users should understand: The authors acknowledge that SAILS is less effective when features are strongly correlated with each other or when three or more features interact together, meaning it works best for simpler interaction scenarios and may miss important dynamics in complex real-world data.

Validation through both controlled tests and real applications demonstrates practical value: The empirical validation using both simulations (where ground truth is known) and real-world tasks shows the method can reliably identify and characterize interactions, supporting its use in responsible AI auditing and model governance.