We study how to choose product assortments directly from transaction data without first estimating a choice model. The framework extracts partial preference orders from observed purchases and optimizes worst-case revenue over preferences consistent with those observations. We develop exact and approximation algorithms with performance guarantees and evaluate when this approach is especially useful, including settings with limited data or model misspecification.

Generative AI allows platforms to choose ad content as well as ad positions and prices. We study how these decisions interact in sponsored search. Targeting the top ad can redirect unmatched consumers to lower positions and increase total revenue, even when it reduces clicks at the top. Content design also changes advertisers’ incentives and can make it optimal to place a lower-margin advertiser in the top position.

We study assortment optimization under flexible choice models and downward-feasible constraints. Starting from the independent-demand model, we adjust product revenues to account for substitution effects while keeping the optimization simple. Under suitable conditions on the choice model and constraints, the resulting heuristic requires only two linear programs and achieves a tight approximation guarantee. Computational experiments evaluate its performance across choice models and on real transaction data.

We study online decisions in which items arrive sequentially, admission decisions are irreversible, and the final payoff depends on a stochastic choice among admitted items. The decision-maker also chooses actions such as prices, creating interactions that standard prophet-inequality models do not capture. We develop simple threshold policies with constant-factor performance guarantees under structural conditions on the choice model and examine extensions with admission and resource constraints.

When consumers rely on seller-controlled AI for product advice, does greater trust make them better off? We model a seller’s pricing and disclosure decisions when some consumers trust its disclosure commitment and others are skeptical. The effect of trust depends on the cost of suppressing unfavorable information: trust can reduce consumer surplus when suppression is cheap, improve it at intermediate costs, and become neutral when suppression is expensive. These results show why transparency policies need to account for sellers’ disclosure incentives.

We model a multi-airport system as a multiclass queuing network to predict aircraft travel times. Historical flight trajectories identify traffic-flow patterns and provide estimates of routing and time-varying airport and airspace capacity. Simulations capture interactions among airports and terminal airspace and are evaluated using data from the Guangdong–Hong Kong–Macao Greater Bay Area.