Skip to main content

Could an algorithm help find the right place to resettle refugees?

Could an algorithm help find the right place to resettle refugees?


Researchers are experimenting with machine learning to find the best place for refugees to get a job

Share this story

Day Of Service Honoring Martin Luther King Jr. In Boston
Volunteers paint flags of the world for immigrant and refugee families learning English on Martin Luther King Jr. Day.
Photo by Keith Bedford / The Boston Globe via Getty Images

Researchers have created an algorithm that they claim can find the best location to resettle a refugee in any given country, improving an individual’s chances of getting a job by as much as 70 percent.

The algorithm, which has yet to be tested in the real world, is trained on historical data of resettlements. It was created by analyzing biographical information about refugees, including their age, gender, language skills, and country of origin, as well as where they were resettled and whether or not they got a job. For each country, it looks for connections in the data, like what circumstances led to a refugee quickly getting a job? What circumstances led to the opposite? Based on these patterns, it makes its own predictions about where an individual should be sent, and how well they will do there.

As described in a paper published in Science today, the researchers tested the algorithm by taking recent refugee data for the US and Switzerland and asking it to pick optimum locations for each individual. With the US data, the average predicted probability of employment for an individual increased from 25 percent to 50 percent; in Switzerland, it improved by 73 percent.

“Right now, there’s no purposeful matching that happens at all.”

These improvements may sound improbably large, but the researchers say part of the reason for this is that countries generally don’t optimize these decisions at all. In Switzerland, for example, incoming refugees are just assigned randomly to one of the country’s 26 cantons (administrative districts), while in the US, assignments are based solely on capacity.

“Right now, there’s no purposeful matching that happens at all,” Jens Hainmueller, a professor in the Department of Political Science at Stanford University who helped create the algorithm, tells The Verge. “It’s literally just, if you come in a given week and there’s a spot in Denver — you’re going to Denver.”

Hainmueller gives the example of a French-speaking refugee who arrives in Switzerland. Although Switzerland is divided into areas of German-speaking and French-speaking individuals, it’s just as likely that the refugee would be sent to the German-speaking part as the French-speaking part. “Given that the algorithm [takes these things into account] it’s not surprising the system has these improvements,” Hainmueller says.

Of course, it’s hard to trust a system that has never been tested in the real world and that just happens to produce amazingly positive results. (It’s doubly hard when the decisions it’s taking could make or break people’s entire lives.) But the team behind the algorithm is confident it’s picking up the right patterns, and say they tested it by matching its predictions against historical data.

“The historical tests give us some confidence that you will see see gains, but we can only say that with certainty when we are able test it [...] in the real world,” says Hainmueller.

But while optimizing for maximum employment is one way to help refugees integrate into their new home, experts say it doesn’t give the whole picture. Mike Mitchell oversees US resettlement programs for the nonprofit HIAS, and he points out that one of the most important and beneficial factors for refugees is not employment, but social connections.

The algorithm ignores a hugely important factor: social connections

“In Western societies like the UK and US, you learn so much through informal networks, through relationships at a church, synagogue, or through your neighborhood. Those relationships lead to opportunities and they strengthen the decisions you make,” Mitchell tells The Verge. “Long-term integration happens through those types of relationships.”

These social contacts can be fostered through a job too, but often they occur outside the workplace. Unfortunately, countries don’t track these sorts of connections when monitoring refugee integration, so factoring them into an algorithm driven by data is pretty much impossible.

Mitchell adds that while the algorithm seems to perform well when looking at refugees as a group, it might fall down in a big way when trying to settle specific cases. “A lot of times, with cases, there may be a health issues, or some reason you can’t resettle a person in one community versus another,” says Mitchell. “These constraints need to be accounted for.”

Hainmueller and his colleague Kirk Bansak, a PhD candidate at Stanford’s Immigration Policy Lab, agree that this is a weakness of their system. But, they say, it doesn’t stop the software from being useful. Individual cases with special circumstances should just be given extra scrutiny. “The algorithm would be giving recommendations to case officers, and the case officers would make the final determinations,” Bansak tells The Verge. “We feel there should be a human override to make sure nothing strange is going on with the assignment that the algorithm has missed.”

Hainmueller and Bansak note that the algorithm is also cost-efficient (the data it uses is already being gathered by resettlement agencies), and could be integrated relatively easily with existing decision-making processes.

Mitchell is less certain. He says that even if the costs are minimal, making improvements to refugee resettlement processes in the US is stymied by two things: institutional inertia and the political climate. The systems in place now “have not changed much in 40 years,” he says. ”So there’s a resistance, or at the very least a skepticism, of change.”

Political hostility might be the main opposition to tools like this

An algorithm that can show how refugees can contribute to the economic success of a country would also be treated with hostility for political reasons. “A tool like this would validate the value that refugees and immigrants bring to communities. And there are those that, because of a certain philosophy, believe they have to prove otherwise,” says Mitchell. He adds, “There are so many challenges right now, just on a day-to-day level, with the administration trying to stop refugee resettlement.”

Nevertheless, Mitchell is enthusiastic about the algorithm’s potential — as long as it’s watched closely. “A tool can work on improving the decision making, but ultimately it should be informing the decisions, not making them,” he says. “But I think this is a great step, and it shouldn’t be discounted for what it doesn’t have right now.”

Case workers who help settle refugees and asylum seekers have expressed similar thoughts, say Hainmueller and Bansak. It’s not that they think the algorithm is some deus ex machina that’s going to instantly solve the important and difficult problems of refugee resettlement. Instead, it might jump-start some changes in the field, giving a new dimension and life to an area of political policy that has been sidelined and constrained for decades.

“People are excited to get this kind of help from an algorithm to make decisions,” says Hainmueller, “because right now they’re kind of flying blind.”