When you search for a product on an eCommerce platform, request a ride-hailing service, or browse a job board, a matching algorithm determines what you see. Historically, these algorithms had a single objective: maximize utility for the end user. If the user found a relevant product or a cheap ride quickly, the algorithm was deemed successful.
However, modern digital platforms are not single-user utilities; they are multi-sided marketplaces. When an algorithm exclusively optimizes for the consumer, it operates in a vacuum that can inadvertently harm other participants in the ecosystem. Neglecting the supplier side can lead to severe income inequality for gig workers, the systemic exclusion of niche content creators, or anti-competitive environments for third-party sellers.
To prevent these systemic failures, a growing body of research focuses on multi-stakeholder and two-sided fairness. The core intuition is that fairness cannot be evaluated or optimized from a single perspective. Instead, recommendation and matching systems must treat the platform as an interconnected ecosystem where the needs of consumers, providers, and sometimes broader society are balanced simultaneously.
The Multi-Stakeholder Ecosystem
To understand why traditional fairness metrics fail, we must first look at the competing incentives inherent in multi-sided platforms. As categorized by Sahil Verma 0003 et al., a comprehensive fairness framework must account for at least three distinct groups of stakeholders:
- Demand-side stakeholders (Consumers): Users who require accurate, personalized, and non-discriminatory recommendations or services.
- Supply-side stakeholders (Providers): Sellers, gig workers, or creators who require equitable exposure, fair pricing, and sustainable income opportunities.
- Side stakeholders (Platform/Society): Third parties or systemic goals, such as ensuring platform robustness, preventing monopolistic behavior, or promoting sustainable consumption.
┌────────────────────────┐
│ Platform │
│ (Revenue, Robustness) │
└───────────┬────────────┘
│
┌───────────────┴───────────────┐
▼ ▼
┌────────────────────┐ ┌────────────────────┐
│ Demand Side │ │ Supply Side │
│ (Consumers) │◄────────►│ (Providers) │
│ Utility, Fairness │ │ Exposure, Earnings │
└────────────────────┘ └────────────────────┘
When these incentives collide, single-objective optimization breaks down. For example, in eCommerce search, Dietmar Jannach et al. point out that platforms face a constant tension between maximizing immediate customer relevance and ensuring fair exposure for diverse, third-party suppliers. Without explicit multi-stakeholder interventions, ranking algorithms naturally drift toward popularity bias or unfairly prioritize the platform's own private-label products over external competitors.
A similar tension exists in crowdsourcing and gig labor platforms. Ria Mae Borromeo et al. observe that evaluation metrics on these platforms heavily favor requester-centric goals like minimizing cost and maximizing speed, while systematically neglecting worker welfare, satisfaction, and long-term retention.
Why Single-Sided Fairness Fails
A natural first instinct might be to apply traditional algorithmic fairness techniques—such as demographic parity or equalized odds—to the consumer side alone. However, the literature demonstrates that single-sided interventions often exacerbate imbalances on the other side of the market.
In a comprehensive survey of the field, Di Jin et al. emphasize that ignoring provider-side dynamics leads to the systemic exclusion of smaller or niche participants. Even if a system is perfectly fair to its users, it can still starve high-quality, minority creators of the exposure they need to survive, ultimately degrading the diversity of the entire ecosystem.
Furthermore, even advanced causal fairness frameworks fall short if they remain single-sided. As noted by Yaochen Zhu et al., while path-specific counterfactual inference can successfully isolate and remove discriminatory biases affecting users, extending these surgical causal interventions to protect item providers remains an unresolved challenge.
The danger of single-sided optimization becomes highly visible in physical matching markets like ridesharing. As demonstrated by Ashwin Kumar et al., when a platform attempts to fix equity or access issues for only one stakeholder group (such as passengers in specific neighborhoods), it frequently introduces severe income disparities or operational inefficiencies for the drivers on the other side. Because the market is a zero-sum game for immediate resources, optimizing for one side inevitably squeezes the other.
Mechanisms for Joint Optimization
To resolve these tensions, researchers have developed algorithmic mechanisms designed to optimize for multiple competing constraints simultaneously. These solutions generally fall into three methodological paradigms: fair division, game-theoretic pricing, and long-term sequential optimization.
1. Fair Division and Two-Sided Matching
When resources are limited or indivisible, the problem can be framed as a fair division game. For instance, in peer-review systems where papers must be matched to reviewers, Justin Payan et al. introduced mechanisms that guarantee Envy-Free up to one item (EF1) fairness. By greedily optimizing the order in which papers "choose" reviewers, their approach ensures high-quality matches for the authors while preventing any single paper from being unfairly disadvantaged or any reviewer from being overloaded.
In highly competitive matching markets like employment or dating, Nan Li et al. introduced a framework that balances user utility with two novel metrics: envy and inferiority. Their system explicitly prevents users from being recommended items or positions they have virtually no chance of securing, thereby protecting the user's experience while simultaneously satisfying the selective constraints of the providers (e.g., recruiters).
2. Game-Theoretic Pricing and Incentives
Instead of treating fairness as a rigid mathematical constraint, some frameworks use economic levers to align stakeholder interests. In the hospitality industry, Errikos Streviniotis et al. modeled dynamic hotel pricing as a cooperative game. By leveraging Owen values—a game-theoretic concept for distributing payoffs among coalitions—their policy dynamically adjusts room prices based on supply and demand. This ensures that consumers receive equitable pricing while competing hotels maintain balanced exposure and profit opportunities. On the gig economy side, Ashwin Kumar et al. demonstrated that introducing simple, targeted financial incentives can successfully guide driver distribution to satisfy passenger demand without compromising driver wage equity.
3. Long-Term Sequential Optimization
Fairness is rarely a static, one-shot problem; an intervention that seems fair today might cause severe imbalances over the course of a week. This temporal dimension is crucial in ride-hailing platforms. Jeffrey Chan et al. showed that optimizing purely for immediate driver income equity can severely degrade passenger service quality by increasing wait times. To solve this, the authors utilized Multi-Objective Multi-Agent Reinforcement Learning (MOMAQL). By training agents to predict future demand patterns, the system looks ahead to distribute rides in a way that smooths out driver income inequality over a long horizon (e.g., weekly earnings) while maintaining reliable, accessible service quality for passengers across diverse geographic neighborhoods.
The Path Forward
Transitioning from user-centric design to multi-sided fairness requires a fundamental shift in how we build and evaluate algorithms. As summarized by Surya Kallumadi et al., modern ranking and recommendation systems must move beyond simple relevance metrics. The future of the field lies in architectures that treat customers, suppliers, and broader societal values not as independent, conflicting constraints, but as deeply interconnected components of a single, sustainable marketplace.
Key Foundational Work
Facets of Fairness in Search and Recommendation (2020) — contributed 3 concept elements to this direction.
ECOM'20: The SIGIR 2020 Workshop on eCommerce (2020) — contributed 3 concept elements to this direction.
On Benchmarking for Crowdsourcing and Future of Work Platforms (2019) — contributed 1 concept elements to this direction.