TLDR: This paper introduces a “ratio-based Shapley value” for collaborative machine learning, offering an alternative to the traditional additive Shapley value. It measures a participant’s contribution by their relative improvement to a model’s performance rather than absolute gain. This new method is proven to satisfy the same fairness and incentive axioms as the additive approach, making it particularly suitable for scenarios with heterogeneous or redundant data, and highlighting that current fairness axioms allow for multiple valid reward mechanisms.
In the evolving landscape of artificial intelligence, collaborative machine learning has emerged as a powerful paradigm, allowing multiple data owners to pool their resources and jointly train models. This collaboration often leads to improved predictive performance, but it also introduces a significant challenge: how to fairly reward each participant for their contribution and ensure everyone is incentivized to cooperate?
Traditional approaches to reward allocation in cooperative settings often rely on concepts from game theory, such as the Shapley value. This value typically measures a participant’s contribution based on the absolute increase in utility they bring to a coalition. For instance, if a participant improves a model’s accuracy from 90% to 92%, the additive Shapley value would consider this a greater contribution than improving a model from 10% to 20%, because the absolute gain (2%) is higher in the first case.
However, a new research paper titled “A Ratio-Based Shapley Value for Collaborative Machine Learning – Extended Version” by Björn Filter, Ralf Möller, and Özgür Lütfü Özçep from the University of Hamburg, Germany, proposes an alternative perspective. The authors argue that in many collaborative machine learning scenarios, relative improvement might be a more meaningful measure of contribution than absolute gain. Imagine a situation where a participant doubles a weak model’s performance from 10% to 20% accuracy. While the absolute gain is only 10%, the relative impact is substantial. The existing additive Shapley value might undervalue such a contribution compared to a smaller absolute gain on an already strong model.
Introducing the Ratio-Based Shapley Value
To address this, the researchers introduce a novel ratio-based Shapley value. Instead of asking “how much does this player increase utility?”, their approach asks “by what factor does this player improve the coalition’s performance?”. This fundamental shift from additive to multiplicative reasoning allows for a reward scheme that better reflects contextual importance, novelty, and leverage, especially in settings with diverse or redundant data sources.
The new ratio-based value is integrated into an existing incentive-aware model reward framework, which allocates non-monetary model quality as a reward. This framework uses a ρ-scaling mechanism, where ρ (a value between 0 and 1) controls the balance between maximizing overall social welfare and ensuring strict fairness. The participant with the largest relative contribution receives the maximum possible reward.
Maintaining Fairness and Incentives
A crucial aspect of this new approach is that it formally defines the ratio-based value and proves that it satisfies the same set of incentive conditions as the additive formulation. These conditions include adapted versions of fairness, individual rationality (ensuring participants receive at least what they could achieve alone), and stability (ensuring the grand coalition of all participants is stable). This means the ratio-based method offers a mathematically sound alternative without sacrificing desirable theoretical guarantees.
The paper highlights several practical advantages of the ratio-based approach. It is particularly well-suited for scenarios involving heterogeneous data quality, where a small amount of high-quality data can drastically improve a weak model. It also performs better in situations with redundant contributions, where marginal additive gains become small due to overlapping information, and in early-stage model building, where absolute gains are small but relative improvements are large. This makes it more responsive to underrepresented data and can incentivize smaller players who make relatively impactful contributions.
Also Read:
- Adaptive Privacy Budgets and Clipping Improve Federated Learning Performance
- PubSub-VFL: Enhancing Collaborative AI Training Efficiency with a Publisher/Subscriber Architecture
Implications for Cooperative AI
One of the most significant theoretical insights from this research is the demonstration that the existing axiomatic framework for incentive-aware model rewards does not uniquely determine the Shapley value. This means that different, yet equally valid, interpretations of “contribution” (additive versus multiplicative) can exist within the same axiomatic structure. This opens up a non-trivial space of compatible mechanisms for designing collaborative machine learning systems.
The authors suggest that future work could explore characterizing the full set of reward mechanisms that satisfy these axioms, refining or extending the current axiom set to favor specific types of fairness, or investigating hybrid reward schemes. This research significantly broadens the toolkit for designing fair and strategically robust collaborative learning protocols, especially where model access is the reward and contributions are nonlinear and context-dependent. You can read the full paper for more technical details here: A Ratio-Based Shapley Value for Collaborative Machine Learning – Extended Version.
