ExplainReduce: Generating global explanations from many local explanations

Local explanations can be efficiently summarized into global ones: The paper demonstrates that large sets of local explanations (generated by methods like LIME and SHAP) can be reduced to just a handful of representative models—as few as five in many cases—without losing fidelity to the original AI system's behavior.

Simpler explanations may be more practical: By condensing many local explanations into a small "proxy set," ExplainReduce creates more manageable global explanations that could be easier for non-technical stakeholders to understand and audit compared to examining thousands of individual local explanations.

The approach works across different AI systems: As a model-agnostic method, ExplainReduce can be applied to various types of "black-box" machine learning models, making it potentially useful for explaining different AI systems used in practice without requiring access to their internal workings.

Computational efficiency matters for real-world deployment: The paper shows that greedy optimization techniques can efficiently solve the reduction problem, suggesting the method could be practical for organizations wanting to generate global explanations without prohibitive computational costs.

Trade-offs between explanation detail and usability exist: While reducing many explanations to five representative ones improves usability, organizations implementing this approach would need to consider whether this level of summarization adequately captures important variations in how the AI system makes decisions across different scenarios.