Two-Sided Fairness

Focuses on equitable exposure and opportunity for content creators, sellers, or service providers on a platform. The goal is to prevent popular items from monopolizing visibility and ensure new or niche producers have a fair chance to be discovered.

• Exposure fairness: Measures whether different producer groups receive visibility proportional to their merit or relevance.
• Minimum guarantee policies: Hard constraints that ensure every qualifying producer receives a baseline number of impressions over a time window.
• Merit-based allocation: Distributes exposure based on a producer's quality score rather than historical popularity, preventing rich-get-richer dynamics.

Ensures that recommendations and personalization serve all user segments equitably, avoiding systematic degradation of experience quality for any demographic group. This is the traditional domain of algorithmic fairness applied to the demand side.

• Utility parity: Guarantees that recommendation relevance scores are not consistently lower for protected user groups.
• Calibration by group: Ensures predicted engagement probabilities match actual outcomes equally across all consumer segments.
• Representation in training data: Addresses the root cause of consumer-side unfairness by ensuring diverse user behavior is adequately sampled during model training.

A central tension in two-sided systems where optimizing for one stakeholder's fairness metric degrades outcomes for the other side. A platform must explicitly navigate this Pareto frontier to find an acceptable equilibrium.

• Producer fairness vs. consumer relevance: Forcing exposure for lower-quality producers can reduce the average relevance of recommendations seen by consumers.
• Consumer fairness vs. producer efficiency: Over-optimizing for underserved consumer niches may concentrate exposure on a narrow set of producers who cater to those niches.
• Dynamic weighting: Production systems often use tunable parameters that allow platform operators to shift the balance in response to business objectives or regulatory pressure.

The mathematical framework for implementing two-sided fairness, where a single model or ranking function is trained to optimize multiple competing objectives simultaneously using techniques from constrained optimization and Pareto efficiency.

• Scalarization: Combines consumer utility and producer fairness into a single weighted loss function, where the weights control the trade-off.
• Constrained optimization: Maximizes consumer utility subject to hard constraints that producer fairness metrics must exceed a specified threshold.
• Multi-gradient descent: Computes separate gradients for each objective and finds a common descent direction that improves all objectives without sacrificing any single one.

Algorithmic frameworks that explicitly model and distribute the limited resource of user attention across producers. These mechanisms treat visibility as a scarce commodity to be allocated according to fairness principles.

• Position-based exposure models: Quantify the probability a user will examine each ranked position, creating a budget of total expected exposure to distribute.
• Fair division rules: Apply principles from social choice theory, such as proportionality and envy-freeness, to the allocation of recommendation slots.
• Amortized fairness: Ensures fairness constraints are satisfied over a rolling time window rather than in every individual recommendation request, allowing for more efficient short-term optimization.

The dynamic, cyclical effects where today's fairness interventions reshape the data distribution that trains tomorrow's models. Two-sided systems are particularly susceptible to emergent bias from these feedback cycles.

• Producer retention dynamics: Fair exposure for new producers increases their likelihood of remaining on the platform, gradually shifting the producer quality distribution.
• Consumer preference shift: Exposure to diverse producers can broaden consumer tastes over time, reducing the inherent tension between fairness and relevance.
• Strategic manipulation: Producers may alter their behavior to game fairness mechanisms, requiring robust counterfactual evaluation to measure the true causal effect of interventions.