Explaining the root causes of unit-level changes

Current AI explanation methods have a blind spot: Existing explainable AI tools cannot adequately explain why an individual's outcome changed over time—they can explain a single prediction, but not what caused the change between two points in time, which is crucial for understanding real-world decisions affecting people.

The paper proposes a game theory-based solution: The authors develop methods using Shapley values (a concept from game theory) to separate two sources of change: shifts in input values (e.g., someone's education level) versus shifts in the underlying "mechanism" (e.g., how the job market values that education), allowing clearer attribution of what drove the change.

The methods satisfy important fairness principles: The proposed approaches satisfy two key axioms—meaning they avoid double-counting causes and properly allocate credit—which are essential safeguards for ensuring explanations are logically sound and fair when used in high-stakes decisions.

Real-world validation on earnings: In a case study analyzing changes in individual earnings in the US, the methods produced "sensible results," suggesting practical applicability for understanding economic mobility and identifying whether income changes were driven by personal factors or broader economic shifts.

Scalability remains an open question: While simulations tested reliability, the paper indicates scalability concerns, suggesting these methods may face computational challenges when applied to very large datasets or complex real-world systems.