Equity by Design: Fairness-Driven Recommendation in Heterogeneous Two-Sided Markets

Two-sided marketplaces embody heterogeneity in incentives: producers seek exposure while consumers seek relevance, and balancing these competing objectives through constrained optimization is now a standard practice. Yet real platforms face finer-grained complexity: consumers differ in preferences and engagement patterns, producers vary in catalog value and capacity, and business objectives impose additional constraints beyond raw relevance. We formalize two-sided fairness under these realistic conditions, extending prior work from soft single-item allocations to discrete multi-item recommendations. We introduce Conditional Value-at-Risk (CVaR) as a consumer-side objective that compresses group-level utility disparities, and integrate business constraints directly into the optimization. Our experiments reveal that the “free fairness” regime, where producer constraints impose no consumer cost, disappears in multi item settings. Strikingly, moderate fairness constraints can improve business metrics by diversifying exposure away from saturated producers. Scalable solvers match exact solutions at a fraction of the runtime, making fairness-aware allocation practical at scale. These findings reframe fairness not as a tax on platform efficiency but as a lever for sustainable marketplace health.

💡 Research Summary

The paper tackles fairness in modern two‑sided marketplaces where both consumers and producers have heterogeneous incentives and the platform must also satisfy commercial objectives. While prior work has shown that, under very restrictive assumptions (single‑item recommendation, extreme consumer‑to‑producer imbalance), enforcing producer‑side exposure guarantees can be “free” – i.e., without hurting consumer utility, the authors argue that real systems operate far beyond these assumptions. They therefore propose a comprehensive formulation that simultaneously addresses three sources of heterogeneity: (1) multi‑item recommendation lists (k > 1) with discrete binary allocations, (2) group‑level consumer fairness using Conditional Value‑at‑Risk (CVaR) to compress disparities across user segments, and (3) business constraints such as a minimum Gross Merchandise Value (GMV) threshold that ties producer value to platform revenue.

Formally, the allocation matrix w ∈ {0,1}^{m×n} must satisfy ∑j w{ij}=k for every consumer i, while producer exposure is bounded below by a factor γ of the maximum achievable minimum exposure. Consumer utility is measured by relevance loss L_Rel_g for each user group g, and the CVaR objective minimizes τ + (1/(1‑α)G)∑_g max{L_Rel_g − τ, 0}, where α controls risk aversion (α = 0 reduces to mean optimization, α → 1 approaches a max‑min group objective). The GMV constraint ∑j v_j ∑i w{ij} ρ{ij} ≥ θ·V_GMV^max ensures that any fairness intervention does not sacrifice a predefined share of the platform’s revenue potential.

Because the resulting problem is a mixed‑integer program (NP‑hard), the authors develop three scalable solution strategies: (i) linear‑program relaxation followed by a careful rounding scheme, (ii) an augmented Lagrangian (ADMM) approach that treats the fairness and business constraints as penalties, and (iii) a gradient‑based continuous relaxation with a post‑hoc discretization (hill‑climbing). All methods achieve near‑optimal solutions (≤ 2 % optimality gap) while being orders of magnitude faster than off‑the‑shelf MIP solvers, making them viable for production‑scale workloads.

Empirical evaluation uses three public datasets representing music streaming (Spotify playlists), e‑commerce (Amazon product grids), and movie recommendation (MovieLens). Experiments vary k (5 and 10), γ (producer exposure), α (CVaR risk level), and θ (GMV threshold). Key findings include: – The “free fairness” phenomenon disappears as soon as k > 1; enforcing producer exposure reduces average consumer relevance by 10‑30 % depending on the dataset. – Moderate producer fairness (γ ≈ 0.3‑0.5) combined with a high GMV threshold (θ ≥ 0.9) actually improves revenue by 2‑5 % because exposure is diversified away from saturated producers. – Raising α to 0.8‑0.9 compresses group‑level utility gaps by over 40 % while incurring only a modest (≤ 5 %) drop in overall relevance. – CVaR‑based optimization consistently outperforms mean‑utility or max‑min individual‑utility baselines in protecting disadvantaged user groups, confirming the necessity of a group‑level risk measure.

The paper thus re‑examines the theoretical “free fairness” result, showing it is a special case limited to single‑item, highly imbalanced settings. In realistic multi‑item environments, fairness always entails a trade‑off, but the trade‑off can be managed so that platform health (both user satisfaction and revenue) improves. The authors conclude by highlighting the practical relevance of their framework for large‑scale recommender systems and suggest future extensions such as incorporating positional bias, dynamic user behavior models, and intersectional fairness across multiple user attributes.