Sadegh Shirani

Sadegh Shirani grew up listening to his father, a high school economics teacher, discuss the complexities of the economy. What most fully captured Shirani’s attention was not necessarily how much his father knew, but what he didn’t know — what no one could know.

“I could see how complex and uncertain these economic analyses were,” Shirani says. “This uncertainty was intriguing to me: How should we think about an unknown future? Is there any way to think, scientifically, about what might happen next?”

By the time Shirani began his undergraduate studies at Sharif University of Technology in Tehran (double majoring in industrial engineering and math), he’d taken a particular interest in probability theory and how it can address real-world uncertainty, initially gravitating to applications in financial mathematics. He sought every probability-related course he could find at the university, taking courses offered by three different departments.

Just before Shirani graduated, he read some papers on queuing theory by Stanford GSB Professor J. Michael Harrison that instantly transformed Shirani’s aspirations. “I saw a type of math that was similar to what I loved in financial mathematics, but it was being used in broader applications, like in healthcare management, scheduling, and inventory management,” Shirani says. He envisioned his research having a bigger impact — “preferably, a social impact,” he adds.

After earning a master’s in probability at Sorbonne University in Paris, Shirani got his Master of Management Sciences degree at the University of Waterloo in Canada, with a focus on waiting times in hospitals, inspired by Harrison’s theories. So when Shirani considered the final stage of his academic studies, “It was clear to me that I wanted to go to the GSB,” he says.

At Stanford, Shirani met with Harrison to build on his research. Shirani began to see how his background in probability and stochastic modeling — the study of dynamic systems that evolve over time with some level of randomness — could be leveraged to tackle a specific problem with various applications: accounting for uncertainty within networked experiments.

What sparked your interest in wanting your work to have some sort of social impact?

During my undergraduate studies, I spent a year and a half as a junior data analyst trying to find some solutions to saving Lake Urmia in Iran, one of the biggest salt lakes in the world. Because of mismanagement and bad policies, the lake was drying up. I studied the data and tried to think as a part of a big team about how we should model, analyze, and propose solutions to restore this lake.

Our team predicted that if this lake got too low, at least 14 million people’s lives would be directly impacted. This program touched something inside me. By analyzing the data, I could see clearly how poor decisions made by politicians were affecting the lives of future generations.

Now, the way I think about real-world problems ties back to how good or bad decisions cause good or bad outcomes, whether in a queuing setting, a healthcare setting, or in broader settings with a causal perspective.

What were the benefits you hoped to unlock by connecting causal inference and machine learning with your deep background in probability, as suggested by Professor Harrison?

I’ll answer your question with an example. Consider a ridesharing company that runs thousands of experiments each month. They may run experiments on changing their pricing algorithm, for instance. For those kinds of experiments, they select a random subset of the drivers and apply the new feature to those drivers. The other drivers maintain the status quo. The goal is to estimate the causal effect of delivering that new feature to all drivers.

But the drivers are indirectly impacting each other. Maybe some drivers, because of the new feature, are motivated to drive more. The available demand for the other drivers, who didn’t receive the new treatment, might be lower. They’re impacting each other in an indirect way, through this demand mechanism. And as a result, the company may overestimate the effect of this new feature.

We call this the network interference problem — it’s the possibility that if you’re going to deliver a new treatment to some individuals in an experiment, it may spill over and also impact other individuals throughout the network.

Why is the network interference problem such a hard one to solve?

In many instances, this underlying network is very difficult to observe. For example, we cannot see how all of these rideshare drivers impact each other. In other words, we have a fully unknown or partially unknown underlying network.

How did your background help you to tackle that problem in a new way?

My idea was, maybe we can try to track how a networked population evolves in response to an experimental intervention, and this evolution could help us compensate for our lack of knowledge about the network structure itself.

I started to think about similar physical phenomena. My advisor at Stanford, Professor Mohsen Bayati, introduced me to his research papers in statistical physics. Reading them, I noticed the connections between statistical physics and this problem within experiments.

A couple of our recent papers show that by observing the evolution of a network — say, a network of rideshare drivers — we can actually extract enough information over time to use machine learning algorithms to estimate what would happen if we had delivered a new intervention to all drivers or, alternatively, what would happen if we had delivered the new feature to none of them.

So, basically, you were able to translate network-effects knowledge from physics to social networks?

Exactly.

What answers have you come up with to address how these types of experiments, and the interpretation of experimental results, can be improved?

In my current research, we focus on the methodology to develop a toolbox for experiments, which can then be applied to general settings. To showcase the applicability of our tools, we consider different applications: ridesharing, healthcare, and digital interventions.

In all of these cases, the high-level idea is that we’re trying to learn how a network evolves over time.

How do these insights help to improve decision-making?

Imagine we’re starting an experiment now, and it will be running for the next four weeks. If we can learn how the underlying network evolves over those four weeks, then we can consider different scenarios and compare them to pick the best one.

If, in reality, we delivered an intervention to 50% of the individuals in the experiment, we can then also say, what if we had delivered the intervention to all of those individuals? The counterfactual scenario is like another evolution — we didn’t observe it, but we can try to estimate it.

How do you hope that your research can make a positive social impact?

Dealing with the difficulties of running experiments over complex networks is a very common problem: in healthcare, on ridesharing platforms, in social science.

I hope that my research can help develop a new approach to how we can compensate for a lack of knowledge about network structures. I believe this can be done by observing the evolution of a network over time. Perhaps in my future research career, I can help to develop better, more efficient experimental methods — because sometimes the results of these experiments are quite important.

These kinds of experimental results help decision-makers and policymakers ensure that they’re making the correct decisions.