Qualification: Towards Efficient Data Valuation Based on the Shapley Value
Contribution must be measured before value can be allocated. Jia and coauthors formalize this requirement for machine learning data. Their starting point is an economic question: how should revenue generated by a model be allocated among the persons whose data helped create it? The proposed Shapley value assigns each contributor an average marginal contribution across the possible coalitions of data. It therefore links a traceable change in collective utility to an individual share of that utility. The allocation rule supplies more than a ranking. Group rationality distributes the full yield among contributors. Fairness gives equal value to equivalent contributions and no payoff to data that adds no marginal utility. Additivity permits value to be calculated across multiple uses. The authors identify transparency and decentralizability as consequences of decomposing the utility calculation into separately computable shares. They also show why attribution cannot be treated as a minor accounting step. Exact computation is exponential, approximation introduces error or assumptions, and the utility function must be selected before a contribution can be valued. This evidence qualifies the institutional requirement for machine-to-machine settlement. A runtime can route tasks and record completion without establishing who created the resulting value. Payment, credit, and recurring compensation require an attribution rule bound to a defined outcome metric and to evidence of each participant's marginal contribution. Otherwise the economic decision remains outside the runtime even when execution is fully automated. The article does not study sovereign agent runtimes. Its unit is contributed training data, not autonomous work, recurring service, legal entitlement, or tokenized settlement. It assumes identifiable users, a specified utility function, and access to repeated model evaluations. It does not prove that Shapley allocation is practical, manipulation resistant, or normatively sufficient for every agent economy. The narrower result is still important. Contribution tracing and value allocation are distinct institutional functions. Routing work without both functions leaves the central economic question unresolved.
institutional-designgovernance-designeconomicstokenomicsvalue-attributioncompensationdata-valuationmachine-to-machine-settlement