Equitable Ranking in Heterogeneous Marketplace Ecosystems: A Foundation Model Framework for Quality-Aware Fairness

Abstract

Marketplace ecosystems comprising heterogeneous providers face a fundamental tension between optimizing relevance for established items and enabling equitable participation by under-resourced newcomers. Conventional ranking systems systematically disadvantage items lacking interaction history, reinforcing 'rich-get-richer' dynamics. This paper presents FairRank-LLM, a production-deployed ranking framework that leverages large pre-trained language model (foundation model) semantic representations to simultaneously address fairness, cold-start, and quality objectives. FairRank-LLM introduces: (i) an LLM-based semantic encoder providing zero-shot relevance estimation that eliminates cold-start disadvantages; (ii) a multi-signal scoring function composing semantic relevance, quality, personalization, and freshness signals with learned weights; (iii) a fairness calibration layer enforcing exposure equity through quality-gated probabilistic promotion; and (iv) a distillation and approximate nearest-neighbor pipeline delivering sub-250ms p99 latency at production scale. Evaluated across 960 marketplace items, 2.8M active users, and 47.3M interactions over a 75-day production A/B test, FairRank-LLM achieves a 32.6\% reduction in exposure KL-divergence, a 28.9\% improvement in opportunity gap, and a 63.7\% reduction in cold-start time-to-visibility—all while improving NDCG@10 by 1.3\% and CTR by 4.6\% relative to the production baseline. Provider success rates increase by 19.4 percentage points, demonstrating tangible economic impact. Ablation studies confirm the necessity of each architectural component. Deployment experience surfaces important lessons regarding LLM representation bias, adversarial gaming, and stakeholder alignment, shared here as practical guidelines.