MarS-FL: A Market Share-based Decision Support Framework for Participation in Federated Learning

Xiaohu Wu, Han Yu

Abstract

Federated learning (FL) is rapidly gaining popularity and enables multiple data owners (a.k.a. FL participants) to collaboratively train machine learning models in a privacy-preserving way. A key unaddressed scenario is that these FL participants are in a competitive market, where market shares represent their competitiveness. Although they are interested to enhance the performance of their respective models through FL, market leaders (who are often data owners who can contribute significantly to building high performance FL models) want to avoid losing their market shares by enhancing their competitors' models. Currently, there is no modeling tool to analyze such scenarios and support informed decision-making. In this paper, we bridge this gap by proposing the market share-based decision support framework for participation in FL (MarS-FL). We introduce two notions of $\delta$-stable market and friendliness to measure the viability of FL and the market acceptability of FL. The FL participants' behaviours can then be predicted using game theoretic tools (i.e., their optimal strategies concerning participation in FL). If the market $\delta$-stability is achievable, the final model performance improvement of each FL-PT shall be bounded, which relates to the market conditions of FL applications. We provide tight bounds and quantify the friendliness, $\kappa$, of given market conditions to FL. Experimental results show the viability of FL in a wide range of market conditions. Our results are useful for identifying the market conditions under which collaborative FL model training is viable among competitors, and the requirements that have to be imposed while applying FL under these conditions.

Introduction: The Competitor's Dilemma in Federated Learning

Standard Federated Learning (FL) research typically assumes that participants are either altruistic or can be incentivized through simple monetary rewards. However, a significant real-world hurdle remains: the competitor’s dilemma. In many industries, such as finance or healthcare, potential FL participants are direct rivals in the same market.

Consider a scenario where two banks use FL to improve loan-default prediction. If Bank A has significantly more data than Bank B, the resulting global model might provide Bank B with a massive performance boost, potentially allowing them to offer superior services and capture Bank A’s customers. In such cases, a rational market leader might refuse to collaborate, even if it improves their own model, to avoid eroding their relative competitiveness or market share. This paper focuses on this "Case 3" scenario, where participants are both the model trainers and the eventual users competing for the same customer pool.

Overview of FL architectures and use cases, highlighting the competitive scenario (Case 3) addressed by MarS-FL.

The MarS-FL Framework

To address this dilemma, the authors propose MarS-FL (Market Share-based decision support framework), which bridges the gap between collaborative machine learning and economic market dynamics. The framework relies on three core pillars:

1. Market Dynamics Model

The framework adapts the classic Rust and Zahorik model to characterize customer movement between firms. It accounts for critical factors such as customer loyalty ($r_i$), leaving rates ($v_i$), switching behaviors between competitors, and the impact of new market entrants ($\theta$).

The market dynamics model illustrating customer flows, including loyalty, switching, and new market entrants.

2. $\delta$-Stable Market

A market is defined as $\delta$-stable if no participant’s market share drops by more than a predefined small threshold ($\delta$) after joining the FL ecosystem. This provides a mathematical guarantee of safety for participants concerned about their competitive standing.

3. Friendliness Index ($\kappa$)

The authors introduce a novel metric, the Friendliness Index ($\kappa$), ranging from 0 to 1. If $\kappa \geq 0$, there exists a valid allocation of model performance improvements that keeps the market $\delta$-stable. A higher $\kappa$ indicates an environment more conducive to collaborative learning among rivals.

Key Insights and Game Theoretic Analysis

The interaction is modeled as a non-cooperative game where each participant decides their data contribution level ($x_i$). A surprising result of this analysis is that contributing all available local data is a dominant strategy equilibrium. This implies that firms should not protect their market share by withholding data; instead, the "protection" should occur during the model distribution stage—specifically, how much of the global improvement each firm receives.

Other key insights include:

  • The "Big Player" Burden: Larger firms with higher initial market shares require greater relative improvements in model performance to maintain stability compared to smaller firms.
  • Loyalty as a Buffer: High customer loyalty significantly increases market "friendliness." When customers are unlikely to switch, firms can collaborate with less fear of losing their base.
  • Collective Friendliness: Market friendliness is determined by aggregate factors, such as the total number of non-loyal customers, rather than the characteristics of any single firm.

Experimental Results and Conclusion

Numerical simulations based on various economic parameters (e.g., market growth, loyalty rates) validate the framework. The results show that FL is viable across a wide range of market conditions, particularly in mature markets with high loyalty. The authors successfully derived tight lower bounds ($Q_i^{min}$) for the relative performance improvement required by each participant to remain $\delta$-stable.

While MarS-FL provides a robust theoretical foundation, future work is needed to address parameter estimation uncertainty through robust optimization and to develop practical deep learning training loops that can achieve specific performance allocations. Ultimately, MarS-FL identifies the conditions under which collaborative model training is not just technically possible, but economically viable for competitors.

Summary updated

Abstract

Federated learning (FL) is rapidly gaining popularity and enables multiple data owners ({\em a.k.a.} FL participants) to collaboratively train machine learning models in a privacy-preserving way. A key unaddressed scenario is that these FL participants are in a competitive market, where market shares represent their competitiveness. Although they are interested to enhance the performance of their respective models through FL, market leaders (who are often data owners who can contribute significantly to building high performance FL models) want to avoid losing their market shares by enhancing their competitors' models. Currently, there is no modeling tool to analyze such scenarios and support informed decision-making. In this paper, we bridge this gap by proposing the \underline{mar}ket \underline{s}hare-based decision support framework for participation in \underline{FL} (MarS-FL). We introduce {\em two notions of $\delta$-stable market} and {\em friendliness} to measure the viability of FL and the market acceptability of FL. The FL participants' behaviours can then be predicted using game theoretic tools (i.e., their optimal strategies concerning participation in FL). If the market $\delta$-stability is achievable, the final model performance improvement of each FL-PT shall be bounded, which relates to the market conditions of FL applications. We provide tight bounds and quantify the friendliness, $\kappa$, of given market conditions to FL. Experimental results show the viability of FL in a wide range of market conditions. Our results are useful for identifying the market conditions under which collaborative FL model training is viable among competitors, and the requirements that have to be imposed while applying FL under these conditions.
BibTeX
@article{wu2021mars,
  author = {Xiaohu Wu and Han Yu},
  title = {{MarS-FL: A Market Share-based Decision Support Framework for Participation in Federated Learning}},
  journal = {IEEE Trans. Big Data},
  year = {2021},
}
Figure