Higher-Order Feature Attribution: Bridging Statistics, Explainable AI, and Topological Signal Processing

Feature interactions matter for explainability: The paper addresses a critical gap in current AI explanation methods—most existing approaches assume features work independently, but real-world models often rely on feature interactions (like when two inputs together drive a prediction). This means standard explanation methods may give incomplete or misleading pictures of how AI systems actually make decisions.

Higher-order attribution provides a more complete picture: By extending Integrated Gradients to capture interactions between features, this framework allows practitioners to identify not just which individual features matter, but also how features work together to influence predictions—important for understanding complex model behavior in high-stakes applications.

Mathematical rigor strengthens trustworthiness: The paper establishes formal theoretical foundations connecting feature attribution to statistics and signal processing, which helps validate that these explanation methods are mathematically sound rather than ad-hoc. This theoretical grounding is essential for responsible AI deployment where stakeholders need confidence in explanation reliability.

Practical implications for model auditing: Understanding feature interactions through higher-order attribution could improve model audits and bias detection, since discriminatory behavior often emerges from how multiple demographic or contextual features combine rather than from single features alone.

Bridges multiple disciplines: By connecting explainable AI to statistics and topological signal processing, this work creates a common language across fields, potentially enabling better collaboration between AI developers, domain experts, and auditors in ensuring models behave responsibly.