Shapley values: Better than counterfactuals

Abstract

The counterfactual impact function often leads to pitfalls when considering scenarios with multiple stakeholders. This inspired the use of the Shapley value function, which assigns impact based on properties including efficiency, symmetry, order indifference, and null-player. The Shapley value effectively addresses many issues present in counterfactual impact, such as double-counting, funging, and leveraging. Despite being computationally complex, the Shapley value can be effectively used through a calculator or approximated using methods from machine learning. In scenarios involving opposing viewpoints, the Shapley value avoids potential adversarial behavior by assigning impact based on its definition rather than considering the timing of actions. Additionally, optimizing for Shapley value has been shown to maximize total cost-effectiveness in cases where agents have constant budgets. While the Shapley value is generally advantageous over counterfactual impact, it’s worth noting that its implementation remains computationally challenging for large datasets. – AI-generated abstract.