CFIRE: A General Method for Combining Local Explanations

CFIRE addresses the "disagreement problem" in AI explainability: The paper acknowledges that no single explanation method consistently works best across different models and datasets. CFIRE solves this by combining multiple local explainers rather than relying on one, making explanations more reliable across diverse situations.

The method produces understandable global rules from complex AI decisions: CFIRE converts local explanations (why an AI made a specific decision) into complete global decision rules that are compact and easy to understand, helping stakeholders grasp how a black-box model operates overall.

Robustness was tested at scale across 700 models: Rather than testing on a handful of cases, the researchers evaluated CFIRE on nearly 700 different black-box models across 14 datasets, demonstrating that the approach works reliably in real-world conditions—an important consideration for responsible AI deployment.

The method accounts for the "Rashomon effect" in explainability: The paper recognizes that multiple valid explanations can exist for the same decision, and their evaluation methodology accounts for this complexity, providing a more honest assessment of explainability methods than typical approaches.

Works with any local explainer, increasing flexibility: Because CFIRE can integrate with different explanation methods, organizations can choose explainers suited to their specific needs while still generating trustworthy global rules, making it adaptable to various responsible AI contexts.