STRIDE: Subset-Free Functional Decomposition for XAI in Tabular Settings

Current AI explanations are oversimplified: Most explainable AI methods reduce complex feature effects to single numbers, answering only "what matters" but failing to reveal "how" different features work together—a critical gap for understanding real-world model behavior.

STRIDE makes interaction analysis practical: By avoiding expensive computational methods like Shapley values (which face exponential costs), STRIDE achieves a 3x speedup while maintaining high accuracy (mean R² of 0.93), making it feasible to explain feature interactions in real applications.

Feature interactions can dramatically change predictions: The paper demonstrates that a single interaction between features can shift model performance by 0.008 R² points on California Housing data, illustrating why understanding "how" features interact—not just "what" matters—is essential for responsible AI deployment.

The method is model-agnostic and theoretically sound: STRIDE works across different model types and is grounded in mathematical theory (orthogonality and convergence guarantees), providing confidence that explanations are reliable rather than ad-hoc approximations.

"Component surgery" enables targeted model auditing: The new diagnostic tool allows practitioners to isolate specific feature interactions and measure their exact impact, supporting more rigorous testing for bias, fairness issues, or unexpected behaviors before deployment.