Equity vs. Equality: Optimizing Ranking Fairness for Tailored Provider Needs

Ranking plays a central role in connecting users and providers in Information Retrieval (IR) systems, making provider-side fairness an important challenge. While recent research has begun to address fairness in ranking, most existing approaches adopt an equality-based perspective, aiming to ensure that providers with similar content receive similar exposure. However, it overlooks the diverse needs of real-world providers, whose utility from ranking may depend not only on exposure but also on outcomes like sales or engagement. Consequently, exposure-based fairness may not accurately capture the true utility perceived by different providers with varying priorities. To this end, we introduce an equity-oriented fairness framework that explicitly models each provider’s preferences over key outcomes such as exposure and sales, thus evaluating whether a ranking algorithm can fulfill these individualized goals while maintaining overall fairness across providers. Based on this framework, we develop EquityRank, a gradient-based algorithm that jointly optimizes user-side effectiveness and provider-side equity. Extensive offline and online simulations demonstrate that EquityRank offers improved trade-offs between effectiveness and fairness and adapts to heterogeneous provider needs.

💡 Research Summary

The paper tackles the problem of ranking fairness in information retrieval and recommendation systems from a new perspective that goes beyond the traditional exposure‑based equality approach. Existing work typically assumes that all providers benefit equally from the amount of exposure their items receive, and therefore fairness is defined as giving similar exposure to similar content. However, real‑world providers have heterogeneous business goals: some prioritize visibility, others care more about sales or conversion rates. The authors argue that fairness should be defined in terms of equity—allocating resources according to each provider’s specific needs—rather than treating all providers identically.

To operationalize this idea, the authors introduce a provider‑need model. For each provider g two gain parameters are defined: vᵍₑ (the expected utility from a user examining an item, i.e., exposure) and vᵍ_b (the expected utility from a purchase). These parameters can be derived from business data such as product price, promotional fees, or brand maturity. Using these gains, the authors formulate an “equity loss” that measures the discrepancy between a provider’s desired weight on exposure versus sales (derived from vᵍₑ and vᵍ_b) and the actual weight realized in a given ranking. The loss is computed as an L2 distance across all providers, and it reduces to the classic exposure‑proportional fairness when vᵍₑ and vᵍ_b are equal for every provider.

Building on this fairness metric, the paper proposes EquityRank, a gradient‑based ranking algorithm that jointly optimizes user‑side effectiveness (measured by DCG/NDCG) and provider‑side equity loss. The ranking is represented as a softmax probability distribution over items, making the objective differentiable. The overall loss function is a weighted sum:

 L = – λ_user · NDCG(π) + λ_eq · EquityLoss(π)

where λ_user and λ_eq control the trade‑off between relevance and equity. Gradients of both terms are computed with respect to the item scores, and stochastic gradient descent updates the model parameters. In an online setting, the algorithm updates relevance estimates and the provider gain parameters in real time using unbiased click‑through and purchase estimators, allowing the system to adapt to changing provider preferences.

The authors evaluate EquityRank on large e‑commerce datasets (e.g., Alibaba, Amazon) in two settings. In the offline experiments, ground‑truth relevance labels are available, and the authors report average NDCG (aNDCG) and the equity loss. In the online simulation, user interactions are generated dynamically, and cumulative NDCG (cNDCG) is measured alongside the running equity loss. Compared with state‑of‑the‑art fairness‑aware methods such as FairCo, PG‑Rank, MMF, and exposure‑proportional baselines, EquityRank achieves modest but consistent improvements in relevance (1–3 % higher NDCG) while reducing equity loss by 15–30 %. Importantly, the algorithm benefits both “exposure‑focused” providers (high vᵍₑ) and “sales‑focused” providers (high vᵍ_b), demonstrating its ability to satisfy heterogeneous provider goals.

The paper acknowledges two main limitations. First, the gain parameters vᵍₑ and vᵍ_b must be estimated beforehand; inaccuracies can affect the fairness outcome, suggesting a need for automated or adaptive estimation techniques. Second, the current model only considers two interaction types (examination and purchase); extending the framework to incorporate other user actions such as likes, shares, or add‑to‑cart would increase realism but also complexity. The authors propose future work on robust parameter learning, multi‑behavior extensions, and scalable distributed implementations of the gradient‑based optimizer.

In summary, this work introduces a novel equity‑oriented fairness framework that explicitly models provider‑specific preferences and integrates them into a differentiable ranking objective. The proposed EquityRank algorithm demonstrates that it is possible to improve user relevance while simultaneously delivering fair outcomes that respect the diverse economic objectives of content providers, marking a significant step toward more sustainable and inclusive IR and recommendation ecosystems.