Data-Driven Optimization in Revenue Management: Pricing, Assortment Planning, and Demand Learning

Abstract

This dissertation studies several problems in revenue management involving dynamic pricing, assortment selection, and their joint optimization, through demand learning. The setting in these problems is that customers’ responses to selling prices and product displays are unknown a priori, and the only information the decision maker can observe is sales data. Data-driven optimizing-while-learning algorithms are developed in this thesis for these problems, and the theoretical performances of the algorithms are established. For each algorithm, it is shown that as sales data accumulate, the average revenue achieved by the algorithm converges to the optimal.

Chapter 2 studies the problem of context-based dynamic pricing of online products, which have low sales. For these products, existing single-product dynamic pricing algorithms do not work well due to insufficient data samples. To address this challenge, we propose pricing policies that concurrently perform clustering over products and set individual pricing decisions on the fly. By clustering data and identifying products that have similar demand patterns, we utilize sales data from products within the same cluster to improve demand estimation for better pricing decisions. We evaluate the algorithms using regret, and the result shows that when product demand functions come from multiple clusters, our algorithms significantly outperform traditional single-product pricing policies. Simulations with both synthetic and real data from Alibaba show that our algorithm performs very well, and a field experiment at Alibaba shows that our algorithm increased the overall revenue by 10.14%.

Chapter 3 investigates an online personalized assortment optimization problem where customers arrive sequentially and make their choices (e.g., click an ad, purchase a product) following the multinomial logit (MNL) model with unknown parameters. We develop several algorithms to tackle this problem where the number of data samples is huge and customers’ data are possibly high dimensional. Theoretical performance for our algorithms in terms of regret are derived, and numerical experiments on a real dataset from Yahoo! on news article recommendation show that our algorithms perform very well compared with benchmarks.

Chapter 4 considers a joint assortment optimization and pricing problem where customers arrive sequentially and make purchasing decisions following the multinomial logit (MNL) choice model. Not knowing the customer choice parameters a priori and subjecting to a display capacity constraint, we dynamically determine the subset of products for display and the selling prices to maximize the expected total revenue over a selling horizon. We design a learning algorithm that balances the trade-off between demand learning and revenue extraction, and evaluate the performance of the algorithm using Bayesian regret. This algorithm uses the method of random sampling to simultaneously learn the demand and maximize the revenue on the fly. An instance-independent upper bound for the Bayesian regret of the algorithm is obtained and numerical results show that it performs very well.