Gah-Yi Ban wins 2021 Best Operations Management Paper Award

Gah-Yi Ban, an Amazon Visiting Academic who last August joined the company's Logistics Research Science team, won the 2021 Best Operations Management Paper in Operations Research Award from the Manufacturing and Service Operations Management Society (M&SOM).

Ban, who is also an associate professor of Decision, Operations, and Information Technologies at the Robert H. Smith Business School, University of Maryland, earned the honor for a paper she co-authored with Cynthia Rudin, professor of computer science, electrical and computer engineering, statistical science, mathematics, and biostatistics and bioinformatics at Duke University.

The award is given to a paper published within the past three years in the journal Operations Research which is “deemed by the M&SOM editorial board to be most deserving for its contribution to the theory and practice of operations management”. The pair won for their paper, "The Big Data Newsvendor: Practical Insights from Machine Learning".

The newsvendor problem

In the paper, the authors “propose solving the ‘big data’ newsvendor problem via distribution-free, one-step machine-learning algorithms that handle high-dimensional feature data and derive finite-sample performance bounds on their out-of-sample costs.” The newsvendor problem, a classic in operations management, examines how to determine optimal capacity and inventory in the face of uncertain demand.

“The newsvendor problem is the very basic unit of any inventory problem, and this paper revisits that problem from a machine learning lens, specifically from the lens of having access to a lot of contextual or feature data,” Ban said.

Ban and Rudin detail two one-step approaches to solving the newsvendor optimization problem. “One of them, the kernel weights optimization approach, happened to be much better,” she said. “We show empirically, and also theoretically, that this one-step approach can perform better than two-step approaches to the problem that had traditionally been used.”

Ban noted the approach she and Rudin took involved merging two previously distinct approaches when it comes to data-driven decision making.

“In the past, an optimization researcher might have thought, ‘Okay, we’ll do some statistical work with the data and that's the first step. Then we do optimization.’ It ended up being a two-step thought process,” she explained. “But in this paper, we show we can start right from the data and construct an optimization model that actually includes the available data in the model itself.”

Apart from inventory optimization, the paper’s conclusions also have practical implications for staffing challenges, a topic addressed directly in the paper.

“We studied a nurse staffing problem based on some data from the UK. Normally you have regular nurses on a shift, but because there's uncertain demand, especially in emergency rooms, you may be short of staff and need to call agency nurses who cost a lot more per hour. We incorporated the nurse staffing example into a newsvendor model. For that particular dataset, we show that our best algorithm, the one-step kernel-weights algorithm, beat the practice benchmark by 24% with statistical significance at the 5% level.”

Ban notes that although what she and Rudin proposed in their paper is well accepted now, it was novel when they first worked on it.

“I think what caught everyone's attention back when the working paper version was first released in 2013 was this idea of integrating big data and prescriptive elements — of what you can do with big data together with optimization.”

Her paper has already led to some bonding with her new coworkers on the Logistics Research Science team.

“One of my colleagues, Chinmoy Mohapatra, is an operations PhD from the University of Texas at Austin. He said he knew about my paper,” Ban recalled. “They taught it in his PhD program. In fact, we are working on a project right now, and he noted we may be able to use our one-step approach for it. What a small world!”