Why Attribution Stability Matters More Than Attribution Accuracy

In the world of regulated AI, focusing on SHAP attribution accuracy isn't the best approach. Instead, σ_SHAP, which measures variance across multiple rotated background samples, provides a more defensible metric. This shift is crucial because it ensures consistent, reliable results that are easier to validate and trust, especially in regulated environments where accountability is key. Understanding this change helps us appreciate the importance of stability over mere accuracy, pushing the field toward more robust and dependable AI models.