On Understanding the Influence of Controllable Factors with a Feature Attribution Algorithm: a Medical Case Study

Distinguishing controllable from uncontrollable factors matters for AI explanations: Traditional feature attribution methods treat all factors equally, but the paper demonstrates that separating controllable factors (like interventions we can change) from uncontrollable ones (like age or genetics) provides more actionable insights for decision-makers in medical settings.

Current AI explanation methods can mislead users about what's actually changeable: The research shows that existing explainability algorithms may misrepresent the consequences of changing feature values, potentially leading healthcare professionals or policymakers to focus on factors they cannot actually control.

The CAFA method enables better policy recommendations: By applying their Controllable fActor Feature Attribution approach to COVID-19 non-pharmaceutical interventions, the authors demonstrate that AI can identify which controllable measures (like social distancing policies) actually drive outcomes, while accounting for uncontrollable background factors.

Responsible AI explanations require domain-specific design: The paper's medical case study illustrates that responsible AI systems need to be thoughtfully designed with input from domain experts who understand which factors are actually controllable in real-world applications, rather than applying one-size-fits-all explanation methods.

Better explanations don't require throwing away data: CAFA achieves more accurate attribution of controllable factors while still using the full dataset for predictions, suggesting that responsible AI doesn't always mean sacrificing predictive performance for interpretability.