Two-step counterfactual generation for OOD examples

OOD detection alone isn't enough for safety-critical systems: The paper highlights that while machine learning models can identify out-of-distribution data, they typically cannot explain why something is flagged as OOD—a significant gap for high-stakes applications like healthcare or autonomous vehicles where users need to understand model decisions.

Counterfactual explanations can bridge the explainability gap: The proposed "OOD counterfactual" method generates modified versions of flagged data points that show how the input would need to change to move between different out-of-distribution categories, providing concrete explanations rather than just binary OOD/in-distribution labels.

The method iteratively reveals decision boundaries: By showing how data points transition between different types of OOD examples, the approach helps users understand what specific features or characteristics make something fall outside the model's training distribution, making the model's reasoning more transparent.

Benchmark testing validates the approach: The authors tested their method on both synthetic and real benchmark datasets using multiple evaluation metrics, suggesting the technique is practical and measurable—important for building trust in AI systems deployed in real-world settings.