PERFEX: Classifier Performance Explanations for Trustworthy AI Systems

PERFEX addresses a critical gap in AI explainability: While existing explanation methods focus on individual predictions, this paper introduces a way to explain when and why a classifier works well or poorly overall—information that helps users decide whether to trust the system in specific situations.

The method uses a "meta tree" to identify performance conditions: PERFEX learns which input conditions lead to high or low error rates, creating compact, understandable rules (e.g., "the system performs poorly when feature X is below threshold Y"). This enables both developers and operational users to understand system limitations.

Practical value for decision-makers: In real-world deployment, this approach helps decision-makers know when not to rely on the system's recommendations—a crucial element of responsible AI that goes beyond simply explaining individual predictions.

Maintains accuracy even with weak base classifiers: The paper demonstrates that PERFEX can reliably identify performance patterns even when the underlying classifier struggles to differentiate between classes, making it useful across different model quality scenarios.

Supports both development and deployment phases: The explanations help developers identify missing features or opportunities to combine models during development, while also providing operational guidance for when systems should be used or bypassed in practice.