Assessing Algorithmic Bias in Two-Sided E-Commerce Marketplaces: Implications of Search, Ranking, and Recommendation Systems

Abstract

Two-sided e-commerce marketplaces increasingly rely on search, ranking, and recommendation systems to allocate attention, match buyers to sellers, and mediate competition among listings. These systems operate under strong feedback, since user interactions both reflect and reshape what is shown, purchased, and reviewed. Concerns about algorithmic bias in this context are not limited to interpersonal fairness; they also involve market power, quality discovery, seller entry incentives, and the distribution of consumer surplus. This paper develops a technical framework for assessing algorithmic bias in marketplace retrieval and recommendation pipelines, emphasizing measurable exposure and welfare outcomes under realistic behavioral and platform constraints. A unified model is presented for query-conditioned ranking, recommender-driven discovery, and blended result pages, allowing bias to be decomposed into data, model, objective, and interface components. The study formalizes group-conditional disparities in exposure, click propensity, conversion, and long-run seller viability, and shows how these disparities can arise even when relevance predictions are well calibrated on average. Statistical testing and uncertainty quantification are integrated with counterfactual evaluation methods that correct for position bias, missing-not-at-random logging, and interference across listings. To address dynamic effects, the paper introduces a numerical modeling layer that treats exposure as a propagating signal over the marketplace graph, enabling spectral diagnostics and PDE-inspired simulation with finite element discretizations. Finally, mitigation is treated as constrained optimization over ranking and recommendation policies, with attention to stability, robustness, and operational governance in production systems.