Abstract
Two-sided marketplaces like Etsy and eBay must balance the competing needs of buyers, sellers, and the platform itself. Traditional search ranking models often focus exclusively on buyer relevance, leading to a "rich-get-richer" dynamic that harms marketplace health. This paper presents a joint optimization framework using Evolutionary Strategies (ES) to tune search rankings for complex, non-differentiable goals such as seller fairness (Gini index) and arbitrage prevention. By treating ranking as a policy optimization problem, the authors demonstrate that platforms can achieve a more equitable distribution of sales without sacrificing the buyer experience.
The Challenge of Two-Sided Marketplaces
In e-commerce search, relevance is only one piece of the puzzle. A search engine serves three distinct masters: buyers seeking specific items, sellers seeking exposure, and the platform aiming for long-term growth. Standard "Learning to Rank" (LTR) models typically optimize for metrics like NDCG (Normalized Discounted Cumulative Gain), which focuses purely on relevance. However, this often results in the Matthew Effect, where a small fraction of top-tier sellers capture the vast majority of traffic and sales.
As shown in Figure 1, the top 5% of sellers (the first vigintile) can account for over 70% of marketplace activity. This concentration squeezes out new shops, reduces variety, and makes the marketplace vulnerable to seller "gaming" or arbitrage—such as manipulating prices and shipping costs to artificially boost rank. These market-level goals are often non-differentiable, making them "invisible" to standard gradient-descent-based algorithms.
Methodology: Joint Optimization via Evolutionary Strategies
To address these market-level issues, the authors propose a framework that moves beyond point-wise or list-wise item scoring. Instead, they treat search ranking as a policy optimization problem using two key innovations:
- Market-Level Metrics as Constraints: The model incorporates the Gini Index to measure wealth inequality and Incentive scores to penalize arbitrage.
- Evolutionary Strategies (ES): The authors utilize ES, a "black-box" optimization technique. ES works by perturbing model parameters, evaluating the resulting "fitness" (a weighted sum of relevance, fairness, and platform health), and iteratively moving toward the most successful configurations. This allows the model to optimize for metrics that are non-continuous or non-differentiable.
Furthermore, the authors introduce Stochastic Value Functions (SVF). By injecting noise into the ranking policy, the model learns to "blend" different sub-policies. This prevents the system from over-fitting to a single "winner-take-all" strategy and improves robustness across diverse queries.
Key Results and Insights
The framework was evaluated using Etsy production search logs, covering 5,000 top queries. The results indicate that the ES-based models successfully navigate the trade-offs between relevance and fairness.
As illustrated by the Lorenz curves in Figure 5, the proposed Stochastic Greedy ES (SG-ES) model significantly improves the cumulative wealth distribution among sellers. By adjusting the weights in the fitness function, the platform can "flatten" the sales distribution (moving the curve closer to the diagonal) while maintaining competitive NDCG scores. This proves that relevance can be treated as a constraint rather than the sole objective; a slightly less "relevant" result that supports a new seller can be more beneficial for long-term marketplace health than showing the same top seller repeatedly.
Conclusion and Future Work
This research demonstrates that joint optimization of competing business metrics is not only possible but essential for healthy two-sided marketplaces. While Evolutionary Strategies are more computationally intensive than standard methods, their ability to handle non-differentiable market-level goals provides a powerful tool for platform designers. Future work will focus on improving the computational efficiency of these policy learners and testing their performance in live A/B environments to bridge the gap between offline simulation and real-time user behavior.
Summary updated