The Surprising Amount of Arbitrariness in Shapley-Value Data Valuation

Abstract

The growing economic importance of data has generated interest in principled methods for data valuation. Particular attention has been given to the Shapley value, a result from cooperative game theory that defines the unique distribution of a game's rewards to contributors subject to specified fairness axioms. By casting a machine learning task as a cooperative game, Shapley-based data valuation purports to equitably attribute model performance to individuals. However, the practical operationalization of this process depends on a wide array of practitioner decisions. Many of these decisions lie outside of the scope of the underlying machine learning task, introducing a potential for arbitrary decision making. The sensitivity of valuation outcomes to these intermediate decisions threatens the desired fairness properties. In light of these surfaced concerns, we evaluate the face-value equitability of Shapley for data valuation.