Personalized Marketplace Policies Balance Competing Objectives

Two-sided marketplaces inherently face the challenge of balancing conflicting interests between distinct user groups. This research presents a framework for personalizing marketplace policies, specifically focusing on free-value thresholds for job listings on a large job marketplace. The goal was to maximize employer-side metrics while safeguarding job seeker engagement, a complex task given varying effects across job segments. Standard uplift modeling proved insufficient due to two main issues: the need for multi-objective optimization to manage cross-side externalities, and severe experimental constraints requiring cluster-level randomization with limited discrete treatment levels. The proposed integrated framework tackles these by employing three key components. First, ensemble-based hybrid ranking models optimize target and guardrail metrics separately, reducing guardrail risk by over 10% for equivalent target gains. Second, a treatment effect extrapolation method extends estimates from limited experimental data to untested policy levels, validated by empirical monotonicity assumptions. Finally, post-launch data confirmed the accuracy of extrapolation and guardrail compliance in a production environment, demonstrating that principled methodology can enable effective personalization even under severe experimental and objective constraints.