Using Data Science to Get Food and Volunteers Where They’re Needed Most

More than a quarter of Americans participate in formal volunteer programs, creating more than $160 billion in economic value annually. Many of them connect with volunteer opportunities through VolunteerMatch, an online tool that introduces potential volunteers to nearby nonprofits. Since the pandemic, many of those volunteers have found their way to community organizations, including food banks that are part of Feeding America, one of the largest hunger relief programs in the country.

VolunteerMatch and Feeding America use online platforms that employ algorithms to place volunteers and distribute food. Both try to be both equitable and efficient, serving their partners fairly while making sure that neither volunteer opportunities nor food goes to waste. Yet under real-life constraints, it doesn’t always work out that way. Some organizations are inundated with potential volunteers, while others struggle to find people, and larger food banks may have an easier time securing donations than smaller ones.

Daniela Saban, an associate professor of operations, information, and technology at Stanford Graduate School of Business who studies online marketplaces, has been working with both VolunteerMatch and Feeding America to improve their systems. With Vahideh Manshadi at the Yale School of Management, Scott Rodilitz of the Anderson School of Management at UCLA, and Akshaya Suresh of the RAND Corporation, she created a new volunteer-matching algorithm for VolunteerMatch. With Manshadi and Soonbong Lee, also of Yale, she overhauled Feeding America’s system for managing ad hoc donations. Now she and her colleagues have written a pair of papers describing their efforts.

“Most nonprofits are amazing organizations with very smart people doing very important work, but many of their data science teams are understaffed and overworked,” Saban explains.

The interface of VolunteerMatch, which has since merged with Idealist, was like a job board: Users would enter their zip code and they would be shown a list of volunteer roles they could apply for, arranged in order of proximity and how recently they’d been added. The problem was, all users in a certain zip code would see the same opportunities in the same order, so they were more likely to apply to the opportunities at the top of the list. That left some organizations with more potential volunteers than they could handle while others had none. And, Saban notes, “If there’s an opportunity that gets 50 signups but needs just one volunteer, then there are 49 people who are going to be disappointed.” Those unmatched people not only could have been used to fulfill other opportunities, but they may also be less likely to apply for volunteer opportunities in the future.

The researchers wanted to create a better balance of equity and efficiency. They applied a penalty system to the algorithm: As soon as an opportunity received some interest from users, it was moved further down the list, bumping less popular jobs to the top.

Saban’s team conducted two large-scale field experiments with the new algorithm, which they called SmartSort, in Dallas-Fort Worth and Southern California. In both areas, after the implementation of SmartSort, VolunteerMarch saw an 8% increase in opportunities that received at least one response and no significant decrease in overall matches between organizations and volunteers. “We were able to distribute the sign-ups more fairly, covering the needs of more opportunities, without decreasing the total number of sign-ups,” Saban explains.

VolunteerMatch’s leadership was pleased with the results and has adopted SmartSort as its default algorithm.

Real-Time Results

The Feeding America overhaul presented different complications due to the nature of food rescue. Some donations come in on a regular schedule, which makes it easy to plan distribution, but others are ad hoc, based on what the donors have been unable to sell. The goal of Feeding America is to save as much food as possible while making sure its clients receive an equitable share of donations, but maintaining that balance isn’t always easy.

The ad hoc donations system was powered by an online platform called MealConnect, and it worked fairly well. But Saban and the team saw ways to improve it. Nearly a quarter of the ad hoc donations by total count were going to waste, and despite Feeding America’s efforts to distribute donations fairly, some agencies were receiving more than others. Saban wondered if she could use her expertise to make the system more efficient and equitable.

Saban and her colleagues began their work with Feeding America by interviewing people at the organization and the food banks it serves. The distribution process was relatively straightforward. When an unexpected donation came in, a module called Real-Time matched it with a food bank, which in turn was tasked to offer it to one of its affiliated agencies (such as food pantries and local churches) via email. To facilitate this process, Real-Time presented a ranked list of agencies based on a recommendation system that accounted for factors such as proximity and responsiveness. Each agency had 90 minutes to respond. If it didn’t, Real-Time would offer the donation to the next agency on the list. After four rejections, the donation was considered “lost” and there were no further attempts to match it.

The researchers discovered that smaller agencies often passed on donations not because they didn’t want them, but because no one was available to promptly check Real-Time or they lacked the resources to collect the food. Feeding America, for its part, wanted to ensure fairness in donations, but in order not to waste food, its system ended up offering donations more frequently to agencies that were more likely to accept them.

The variation in response rates and the unpredictability of ad hoc donations made the overhaul of the Real-Time algorithm tricky. To give the smaller agencies more of a chance to get donations, Saban and her team adjusted Real-Time’s algorithm. Each agency is associated with a fixed level of need based on its service area. At each stage in the offer process, several agencies receive donation offers simultaneously, although they have less time — 45 minutes — to respond. If multiple agencies respond, the algorithm uses a rank-based tie-breaking rule to determine who receives the donation.

The researchers built a simulator to test how their algorithm would work in different conditions and areas around the country. They found that the amount of lost donations decreased substantially even as equity among agencies improved. When they presented their findings to Feeding America’s leadership, some of the recommendations, such as the shorter time between offers, were adopted. “We take it as a win,” Saban says.

Saban thinks that other nonprofits that match people and services can learn from the results of these two projects. “In general,” she says, “we tend to think that there’s a tradeoff between equity and efficiency. But sometimes that’s not necessarily the case. There can be a way of improving equity without much harm to efficiency.”