Research

We study the problem when a firm sets prices for products based on the transaction data, i.e., which product past customers chose from an assortment and what were the historical prices that they observed. Our approach does not impose a model on the distribution of the customers' valuations and only assumes, instead, that purchase choices satisfy incentive-compatible constraints. The individual valuation of each past customer can then be encoded as a polyhedral set, and our approach maximizes the worst-case revenue assuming that new customers' valuations are drawn from the empirical distribution implied by the collection of such polyhedra. We show that the optimal prices in this setting can be approximated at any arbitrary precision by solving a compact mixed-integer linear program. Moreover, we study the single-product case and relate it to the traditional model-based approach. We also design three approximation strategies that are of low computational complexity and interpretable. Comprehensive numerical studies based on synthetic and real data suggest that our pricing approach is uniquely beneficial when the historical data has a limited size or is susceptible to model misspecification.

We propose deep neural networks for data-driven stochastic optimization. Using historical data (covariates, decisions, costs), we propose to train a neural network to predict the objective value as a function of both the decision and covariate. After training, for a given covariate, this predicted objective is optimized over the decision variables using gradient-based methods with analytical gradients and Hessians. Performance is characterized by neural network generalization bounds. Comprehensive experiments on newsvendor, personalized assortment pricing, and call center staffing problems demonstrate our method’s strength over existing approaches such as conditional stochastic optimization and analytical approximations, especially when: (1) the objective function is unknown, (2) moderate to large datasets are available, or (3) the problem structure resists simple parametric approximations.  

We study the problem an office meal delivery platform faces every day. Such platforms connect client firms, with catering needs for their employees, to restaurants, by offering a menu of compatible restaurants to all the employees in an office. The platform's revenue is a fixed percentage of the value of the restaurant's orders. The platform incurs the delivery cost. Each restaurant allocates a fixed capacity to the platform, regarding how many meals it can prepare on a given day. We model the problem as a capacitated joint assortment optimization on a bipartite graph with supply and demand nodes. We prove the problem is APX-Hard while its special cases with only one restaurant or one client are NP-hard. Using a Linear Programming relaxation of a reformulation of the original model, and by leveraging the rather soft nature of the capacity constraints in practice, we devise an asymptotically optimal assortment sampling algorithm that may breach any capacity constraint only by a negligible probability, as the number of clients and the restaurant capacities grow. In collaboration with Canada's largest office meal delivery platform, we test the performance of our LP-based sampling algorithm in a carefully controlled field experiment. Our results suggest that our methodology improves the platform's per-employee profit and revenue by at least 14%, and 10%, respectively. 

Assortment optimization under complex customer choice models and operational constraints is a central challenge in revenue management. This is because its non-linear objective function, coupled with large-scale and discrete decision variables, renders it computationally expensive to solve. Meanwhile, first-order methods like gradient descent have seen widespread adoption for continuous optimization in large-scale AI systems. We propose a computational framework that combines first-order methods and neural networks to efficiently solve assortment optimization. Our framework features straight-through estimators, which enable gradients to flow through discrete variables, and utilizes neural networks to perturb the gradient updates. We theoretically ground our framework by proving that our method is guaranteed to converge to the globally optimal solution for the unconstrained problem under the Multinomial Logit model (MNL). Furthermore, recognizing the need for standardized evaluation in this domain, we develop and release a public benchmark dataset, available at https://github.com/wch444/Assortment-Benchmark. This dataset, comprising several challenging assortment optimization problems, serves both to empirically test our proposed framework and to provide a robust testbed for the wider research community to evaluate novel algorithmic solutions.

Piece-rate compensation schemes, where workers are paid for each completed task regardless of the time spent on it, are common in practice. Detecting a potential negative impact on firm performance associated with their use adds to the literature on the challenges of piece-rate compensation schemes. We study a radiology workflow platform that connects off-site radiologists with hospitals. These radiologists select tasks from a common pool, and the service level is characterized by meeting priority-specific turnaround time targets. However, imbalances between pay and workload of different tasks could result in higher priority tasks with low pay relative to workload receiving poorer service than low priority tasks. Using a large dataset, we investigate whether low-priority tasks with a high pay-to-workload ratio have a shorter turnaround time. Then, using the same approach, we investigate whether having many low-priority tasks with high pay-to-workload increases the turnaround time and probability of delay of higher priority tasks. We show that turnaround time is decreasing in pay-to-workload for lower priority tasks, whereas it is increasing in workload for high priority tasks. More importantly, we find evidence of a spillover effect: Having many economically attractive tasks with low priority can lead to longer turnaround times for higher priority tasks, increasing the likelihood that those tasks are delayed. Our results suggest that organizations, where workers have task discretion from a common pool, need to carefully align their piece-rate compensation scheme with the workload of each task. Imbalances may lead to a degradation in the system service level provided to time-sensitive customers.