Introduction
When multiple organizations—such as different hospitals or financial institutions—pool their datasets to train a shared machine learning model, a critical question arises: How much did each dataset contribute to the final model? Accurately answering this is essential for fair profit-sharing, assigning credit, and incentivizing high-quality data contributions.
In cooperative game theory, the Shapley value is the gold standard for measuring these contributions. In machine learning, contributions are typically evaluated based on model performance on a shared hold-out test set. However, agreeing on a single, representative test set is often infeasible, especially if the datasets originate from slightly different distributions or if the models are generative.
A more robust alternative is a Bayesian approach: players agree on a common prior, and a dataset's value is measured by the information gain (the KL divergence) from the prior to the joint posterior. Unfortunately, computing exact Shapley values in this Bayesian setting is computationally brutal. Furthermore, the asymptotic properties of these values as datasets grow infinitely large were previously unexplored. This paper bridges that gap by establishing the convergence properties of the Shapley value in Bayesian learning games and introducing a practical framework for fair data sharing.
Methods: A Parametric Bayesian Learning Game
The authors formalize a parametric Bayesian learning game, where the "characteristic function"—the value of a coalition of datasets—is defined as the posterior-prior KL divergence.
To analyze this theoretically, they utilize a generalized Bernstein-von Mises theorem. This demonstrates that as the number of data points increases, the joint posterior converges to a normal distribution centered at the Maximum Likelihood Estimate (MLE).
The most practical contribution of the paper is a novel online collaborative learning framework. Instead of computing the exact Shapley value at every step (which requires expensive Bayesian inference for every possible combination of players), the framework uses a sample approximation of the Fisher information of each player's data. For a two-player game, the framework dynamically adjusts the rate at which it accepts data from each player so that the determinants of their Fisher information match. By doing this, the system naturally drives the difference in their Shapley values to zero, achieving an asymptotically "Shapley-fair" collaboration.
Key Insights
- The Prior Washes Out: The authors mathematically prove that in the presence of a large quantity of data, the choice of the Bayesian prior becomes inconsequential. The relative contributions of the players depend entirely on the data itself.
- Fisher Information is All You Need: As data scales to infinity, the characteristic function (KL divergence) diverges. However, the difference in Shapley values between any two players converges to a limiting game defined solely by the log-determinant of the joint Fisher information.
- A Massive Computational Shortcut: Because asymptotic Shapley values depend only on Fisher information, we can completely bypass the need to compute complex posterior-prior KL divergences. A consistent estimator of the Fisher information is sufficient to balance contributions and ensure fairness.
Key Results
The authors evaluate their framework on several real-world datasets for Bayesian linear regression and latent space mean estimation, including California Housing, King County House Sales, FaceA (age estimation), and MNIST.
- Theoretical Guarantee: The paper rigorously proves that for any two players, the difference in their Shapley values converges in probability to the difference in a limiting game based on Fisher information.
- Empirical Fairness: In two-player experiments, the proposed online framework successfully forced the Shapley values of both players to converge to the same value over time, ensuring a fair collaboration.
- Quality vs. Quantity: The framework correctly identifies data quality differences. If Player 2 provides noisier data or data with missing values, the framework automatically forces Player 2 to provide a higher volume of data to match the Fisher information (and thus the Shapley value) of Player 1.
- Multi-player Viability: While the core theorem is proven for two players, empirical experiments with three and four players showed that the framework can still successfully drive Shapley values to converge, albeit with more fluctuations in complex environments like VAE latent spaces.
Limitations and Future Work
Currently, the strict theoretical guarantee for the online collaborative framework is limited to the two-player setting. Extending these proofs to an arbitrary number of players is a primary direction for future work. Additionally, the mathematical proofs rely on the assumption that players' data points are independent; relaxing this to account for correlated data across organizations would make the theory more robust. Finally, the authors suggest exploring scenarios where players must collaborate to find the true parameter (e.g., observing disjoint subsets of features), which would provide an even stronger motivation for data sharing.
Summary updated