Berk Ustun is waiting for me outside my office. He has the easy demeanor, the affable, open quality of someone who is comfortable in a number of contexts. Later, as our conversation ranges from optimization problems to basketball to the X-Files, my initial sense is confirmed. Ustun – in the way of CRCS postdocs – is a well-rounded guy. He tells me that, as an undergraduate student, he studied operations research. He learned to translate decision-making problems in the real-world into optimization problems, in which there is an objective to optimize and a set of constraints. There are myriad instances in which it is useful to model decision-making problems as mathematical optimization problems. For instance, Ustun offers, there is the question of how much product Amazon should stock in their inventory to sell to customers in any given month. Optimization techniques allow researchers to formulate problems like this – which on their surface appear nebulous and subject to a cluster of variables – as relatively simple mathematical equations.
This mathematical approach to decision-making was popularized during the Second World War as a means of solving and simplifying real-world problems and making military operations more efficient. As computers became more powerful and data more ubiquitous, it became feasible to apply operations research to a greater number of fields. When he entered his Master’s program, Ustun was interested in decision-making problems related to power systems. Specifically, the question of what type of power plants we should build in order to efficiently meet energy demand. As his program progressed he began using optimization to create simple predictive models in a broader range of fields. He worked on the computational side of things, addressing the question of how, once we formulate optimization problems, we can learn to solve them efficiently. As a Ph.D. student, Ustun began developing systems and risk scores around complicated societal issues like criminal recidivism and disease risk.
These models – in which numbers are assigned to variables that predict outcomes – have been in existence in one form or another since 1920. The process for building such a model is threefold. First you define and state the problem formally. Then you come up with an algorithmic solution. Finally, you apply that solution to different research areas. With a staggering amount of data now available to researchers, this method has become more difficult and more crucial. From a machine-learning perspective, predictive models are very simple. They require whole number coefficients. Where many machine learning methods take computational shortcuts and optimize an approximation for accuracy. Ustun’s simple predictive models optimize accuracy directly. This makes them less scalable but highly accurate. They are less subject to individual and social bias because they take the interactions between variables into account. They only include variables that directly improve accuracy and account for the fact that some variables possess more predictive magnitude than others. They are also accessible to non-technical experts. For example, when calculating the recidivism rate of someone recently released from prison, a judge can solve a simple optimization problem by hand. At Massachusetts General Hospital, this type of predictive model is being used by doctors to predict obstructive sleep apnea. A simple predictive model is also the basis of a diagnostic tool for adult ADD. Ustun’s method customizes models to the problems they are designed to solve. He says, “When you’re able to create simple predictive models that other people can understand, then you can elevate the discussion around the topic in question.”
The elevation of the discussion around data-driven decision-making is Ustun’s primary motivation. His goal is not just to create data-driven decision-making tools and release them into the ether. Rather, he intends to create tools that real people can use to solve real problems, whether by preemptively addressing negative health outcomes or improving the accuracy of recidivism predictions to rectify injustices in the carceral system.
Ustun stresses that while it can take a mere day to formulate an optimization problem, it can take a year to come up with a good method to solve it. He feels most gratified when he has worked out a functional solution to an optimization problem and written it into the software. At this point, its application to real-world contexts becomes tangible. Ustun finds it immensely rewarding to convince machine-learning researchers to create models that bridge rather than widen the disciplinary divide. He says, “There is a lot of discussion around whether or not we should trust models that we don’t understand. In a lot of machine-learning, the models are too complicated for non-technical experts. People need to understand the models and have the ability to validate them and override them. Then society can make data-driven decisions in a way that is better for all of us.”
It is this desire to bridge the disciplinary divide that has brought Ustun to CRCS. He says, “I was drawn to the fact that people here care about doing good theoretical and methodological work. They care about things like fairness and privacy, about the societal issues that optimization methods are designed to solve. It’s rare to get a community of researchers who are really good from a technical standpoint and also care about the problems and see how important they are.” Ustun hopes that the pool of knowledge at CRCS will help him produce ever more nuanced data-driven models in areas with which he is not yet familiar.
Perhaps it’s silly to ask someone who is in the process of defending their thesis what their hobbies are, but Ustun smiles good-naturedly and proffers his love of basketball – playing it and following it. He also loves to travel and recently returned from a trip to China.
“What is it they say,” he jokes. “Spend one third of your time making your parents proud, one third of your time relaxing on a beach, and…” We wrack our brains for the other third, to no avail. Luckily for us, it looks like Ustun will be spending the rest of his time at CRCS.