March 22, 2019

Announcing the winners of the Facebook Mechanism Design for Social Good research awards

By: Eric Sodomka

Last June, at the 19th ACM Conference on Economics and Computation (EC 2018), we introduced the Facebook research awards in mechanism design for social good.

We asked researchers to consider the following problem: Suppose there is an existing online platform that is actively used by the population, and an existing set of social ills (e.g., unemployment, disease, poverty, divisiveness, loneliness). How should one design mechanisms on top of such an online platform to build community in a way that alleviates those social ills?

We received 58 submissions for this award. Amongst those, we chose three winners to each receive an unrestricted gift of $50,000. The selection committee was composed of dozens of people across the company, each with some combination of technical expertise in mechanism design and domain expertise in an area of social good. We announced the winners of these awards at Data Science Africa 2018 in Abuja, Nigeria. In this post, we are pleased to announce these winners to a broader audience and provide an overview of the winning proposals.

The winners are as follows:

  • Mechanisms for Crowdsourcing with Small-Holder Farmers. PI: Mutembesa Daniel (Makerere University). Collaborators: Boi Faltings (EPFL); Christopher Omongo (National Crops Resources and Research Institute); Humphrey Mutaasa (Uganda National Farmers Federation).
  • Modern Social Choice: Mechanisms and Platforms for Large Scale Deliberation. Co-PIs: Ashish Goel (Stanford University); James S. Fishkin (Stanford University); Kamesh Munagala (Duke University).
  • Promoting Diversity in Peer Production through Mechanism Design. Co-PIs: Zhiwei Steven Wu (University of Minnesota); Haiyi Zhu (University of Minnesota).

Mechanisms for crowdsourcing with small-holder farmers

The problem: Farmers in the developing world rely on a healthy crop to provide for their families, but that crop is continually at risk of being destroyed. Diseases and pests can ruin a farmer’s entire harvest, and outbreaks can affect the broader farming community. Such dangers are hard to detect in their early stages, and so farmers rely on agricultural experts, who travel across the country, identifying damaged crops and recommending ways for farmers to mitigate their losses. While farmers benefit from these experts’ insights, experts are limited in number and their coverage is restricted. A farmer may lose precious time while waiting for an expert to arrive, or may not receive any such guidance at all, which could mean the difference in whether or not the farmer’s crop survives.

In their previous work, the PIs gave farmers phones with high-resolution cameras, incentivized the farmers to photograph their crops, and then used those images to identify diseases and outbreaks via machine learning methods. The effectiveness of such an approach requires that pictures are of sufficiently high quality and are taken in relevant locations. The PIs will use this award to scale up their study and run randomized control trials to understand how different incentive schemes affect data quality.

Modern social choice: Mechanisms and platforms for large scale deliberation

The problem: Consider a community that needs to make a public policy decision, such as how to allocate a budget to fund different public programs. In making such a decision, some challenges arise. First, discussions about public policy are often divisive and polarized, with participants resorting to name-calling rather than considering reasoned arguments from the other side. Additionally, the space of possible decisions may be high-dimensional, infinite and lacking known structure in advance. In such a case, it is difficult or impossible for participants to express their preferences fully, and it is similarly difficult to aggregate such preferences to reach a single outcome.

In their previous work, the PIs have taken various approaches to address such challenges, most notably through deliberative polling and the Stanford Participatory Budgeting Platform. Deliberative polling estimates public opinion in a hypothetical world in which the public spends sufficient time communicating and reasoning through opposing arguments and trade-offs. The process involves sampling a set of participants, guiding them through moderated small-group discussions and presenting results on how participants’ opinions changed. Deliberative polls have been run over one hundred times in nearly thirty counties. Similarly, the Stanford Participatory Budgeting Platform has been used to run over 40 participatory budgeting elections and has driven the PIs’ theoretical work in social choice for such complex settings.

The PIs propose to take their past work in deliberative polling and scale it up to more people per poll, more polls at once and with deliberation taking place in a virtual environment. Scaling such a process introduces new challenges in moderating conversations at scaleā€”for example, how one can design the system to encourage civility. The PIs plan to experiment with a variety of approaches, from NLP and machine learning-based approaches that automatically detect uncivil behavior, to interfaces that allow participants to report such behavior.

In addition to this main thread of their proposal, the PIs plan to better understand how to build up a representative sample of the population through ad targeting, how to allow participants to express preferences in complex settings and how to aggregate those preferences.

Promoting diversity in peer production through mechanism design

The problem: In many online platforms, the producers and consumers of content are not representative of each other demographically. For example, in Wikipedia and open street maps, significantly more men contribute to content creation than women. This results in a lack of diversity in content; for example, a relative lack of articles about accomplished women.

The PIs aim to create a more diverse community of content producers. To find a relevant producer, they will consider a system that identifies underrepresented topics and then searches for experts to contribute to those topics through a social referrals mechanism. They will consider badge design to incentivize quantity and quality of content production, considering the tradeoffs between growing the community in the short term and avoiding burnout for more sustainable growth. They plan to deploy their mechanisms to groups within the Wikipedia community and measure the degree to which their mechanisms reduce the gender gap.

Final thoughts

The selection committee was extremely impressed by the large number of top-quality submissions received. While we could only fund three submissions, many more of the submissions contained exciting research directions. We reviewed promising proposals in domains such as refugee resettlement, blood donations, healthcare, civic engagement, charitable donations, misinformation, innovation contests, disaster evacuation and relief, agricultural markets, dating markets, general matching markets and many more. We appreciate the time and effort so many spent writing proposals, and wish all submission authors good luck in pursuing these research agendas.

We especially want to call out the Mechanism Design for Social Good (MD4SG) organization, which is an independent group but was the inspiration for this award. The MD4SG organization has an upcoming workshop, as well as a regular colloquium series and working groups. See the MD4SG website for more information.

Facebook has a number of teams working in social good. If you have expertise in machine learning, AI or economics and computation, and are interested in applying that expertise to social good causes, please apply to work with us. We are hiring interns, postdocs, full-time employees and academic contractors in offices around the world.