Adaptive Two-sided Assortment Optimization: Revenue Maximization

Abstract: We study adaptive two-sided assortment optimization for revenue maximization in choice-based matching platforms. The platform has two sides of agents, an initiating side, and a responding side. The decision-maker sequentially selects agents from the initiating side, shows each an assortment of agents from the responding side, and observes their choices. After processing all initiating agents, the responding agents are shown assortments and make their selections. A match occurs when two agents mutually select each other, generating pair-dependent revenue. Choices follow Multinomial Logit (MNL) models. This setting generalizes prior work focused on maximizing the number of matches under submodular demand assumptions, which do not hold in our revenue-maximization context. Our main contribution is the design of polynomial-time approximation algorithms with constant-factor guarantees. In particular, for general pairwise revenues, we develop a randomized algorithm that achieves a $(\frac{1}{2} - \epsilon)$-approximation in expectation for any $\epsilon > 0$. The algorithm is static and provides guarantees under various agent arrival settings, including fixed order, simultaneous processing, and adaptive selection. When revenues are uniform across all pairs involving any given responding-side agent, the guarantee improves to $(1 - \frac{1}{e} - \epsilon)$. In structural settings where responding-side agents share a common revenue-based ranking, we design a simpler adaptive deterministic algorithm achieving a $\frac{1}{2}$-approximation. Our approach leverages novel linear programming relaxations, correlation gap arguments, and structural properties of the revenue functions.