This Research Award is now closed
At Facebook, we are passionate about connecting people with the businesses they care most about. To achieve the best experience for people, we rely on the application of statistics, machine learning, artificial intelligence, and A/B experimentation to personalize the experience for billions of people across Facebook’s family of apps. One of the areas we’re increasingly focused on at Facebook is how we can design our products and services in ways that preserve privacy while still making powerful statistical inferences from data. With advancements in cryptography and computation, we seek to enable equivalent (and unlock new) functionality while maintaining or enhancing technical guarantees of user-level privacy.
Advancements in the field would facilitate collaborations between industry and academic partners without the risk of private data being learned by the other parties. They would also facilitate learning patterns and trends that can be used to benefit the broader population while still keeping information private for each consumer. Facebook is pleased to invite the academic community to submit research proposals that address these opportunities.
Facebook will grant awards of up to $60K per awardee to fund projects of up to one year in duration. Researchers may submit proposals for privacy-enhancing solutions to multiparty statistical analyses and machine learning problems within advertising. There exists substantial latent demand to facilitate collaborations and to enrich insights while protecting consumer privacy across a wide range of practical circumstances. We encourage proposals of diverse perspectives from emerging scholars and researchers around the world studying real-world applications of cryptography and other privacy technologies.
Proposals should include
- Summary of the project (1–2 pages). Provide a clear explanation of the area of focus, a description of techniques, any relevant prior work, and a timeline with milestones and expected outcomes. Please also include a concise statement of the scientific contribution and routes to eventual deployment. References are not included in the two-page limit, but please keep these concise.
- Curriculum vitae(s). Provide the name of each researcher involved in the proposed work with their CV/résumé or link to Google Scholar page.
- A one-paragraph biography of each principal researcher
- Organization details, including tax information and administrative finance contact details
- Keywords. Please select from the following keywords: advertising algorithms, ad targeting, ad personalization, ad optimization, cross-channel, cross-publisher, audience measurement, sales measurement, cross-publisher, advertising effectiveness, RTB, programmatic ads, ads reporting, experimentation, record linkage.
We are seeking solutions that address the following topics within advertising: analysis of experiments; addressing non-response bias and missing data; complex statistical and machine learning models; causal inference with observational data; record linkage and matching individuals across different channels. Updates to inferences may range from daily to hourly, and may be conducted with a small number of active data-controlling parties or necessitate a distributed approach with billions of individual browsers or devices.
We are particularly interested in applications using one or more of the following privacy-enhancing technologies in their research proposal (we are also open to applications leveraging technologies not listed below):
- Differential privacy
- Secure multiparty computation
- Homomorphic encryption
- Federated learning
- Zero-knowledge proofs
- Secure aggregation
- Trusted execution environments
Below, we describe some specific examples of topics that are of interest to Facebook, but we will gladly review proposals for areas that we have not yet considered:
- A/B experimentation: Questions about effectiveness of ads are best answered by comparing the overall behavior of two groups of users — for example, one group that is exposed to the ads and one group that is not. Computing the difference in purchases or sales across these groups helps advertisers understand the true impact of advertising. Facilitating such user-level testing necessitates identification of users across browsers and devices and also across publishers (which serve the ads) and advertisers (which record sales or purchases). The result, “Lift in sales caused by ads,” is, however, an aggregate measure and is not necessarily private information. The goal is to enable Lift measurement that leverages person-level ad exposure and sales data, while also preserving the privacy of consumer and business data involved.
- Cross-media reach and frequency measurement: Reaching as many people as possible in their target audience is a great way for advertisers to improve awareness of their brand and make sure their brand/product is top of mind for consumers. At the same time, in the interest of increasing efficiency of their media spend, they don’t want to show too many ads to the same person across devices, browsers, apps, and websites. This requires measuring the reach of their advertising campaigns in terms of unique individuals exposed to any ad (across desktop browsers, mobile browsers, mobile apps), as well as the frequency profile of multiple ad exposures to the same individual. Traditional methods have relied on vendors accumulating individual-level advertising exposure data across media channels and then computing reach and frequency. The goal is to enable cross-media measurement that leverages person-level ad exposure data across media while also preserving the privacy of consumers who viewed or interacted with the ad.
- Record linkage: Most advertising optimization and measurement systems are designed to understand the consumers’ immediate purchase actions after interacting with an ad to offer better ad experience for them. Legacy techniques involve facilitating record linkage across businesses (for example, to enable businesses to communicate consumers’ purchase decisions to ad publishers) by sharing encrypted data sets with each other in a secure way, and the record linkage process takes place in one of the parties’ servers. In such a system, privacy guarantees are often contractual. The goal is to use advancements in privacy technology to enable record linkage with built-in technical privacy guarantees.
- Modeled attribution: Advertisers spend their marketing budget across multiple media channels, and consumers can interact with many of these advertisements before they make a purchase on the advertiser’s site. To measure such media investments, advertisers typically use modeled attribution solutions (like multitouch attribution) that incorporate the various ad exposures and purchase behavior in a statistical model to determine relative contribution of certain types of ads toward the likelihood of purchase. The goal is to enable such measurement using privacy-preserving technologies.
- Optimization towards specific objectives: Advertisers tend to pick a measurable objective that they would like to maximize when they buy ads on a specific publisher. As an example, “website conversion” is an objective where advertisers want as many of a particular valuable event (like “add to cart” or “subscribe to newsletter”) to happen as possible. To maximize such objectives for advertisers, publishers train ML models that look for correlation between onsite features (like age or gender a person may have entered into their Facebook profile, or historical click-through rate of that particular ad) and offsite labels (whether or not “website conversion” happened); predictions from these models are then used in delivering subsequent ads to maximize advertiser value. The goal is to enable optimization model training with individual-level features and labels while also preserving privacy of offsite consumer data.
We will prioritize proposals that are pragmatic, use-inspired and grounded in empirical applications. We will pay special attention to proposals that include plans for producing useful, easy-to-use, open source software as part of the research output.
- This RFP is open to current full-time faculty at accredited academic institutions that award research degrees to PhD students. This includes individual researchers addressing a well-defined problem as well as multiple university departments, with a diverse set of skills, collaborating to solve a more complex multidisciplinary challenge.
- Organizations must be an accredited academic institution (PhD-granting) with recognized legal status in their respective country (equal to 501(c)(3) status under the United States Internal Revenue Code).
- Applicants must be the Principal Investigator on any resulting award. Funding will be disbursed to the institution of the Principal Investigator of any resulting award.
- Awards must comply with applicable U.S. and international laws, regulations, and policies.
- Facebook will not provide access to any Facebook, Instagram, or WhatsApp data for awarded projects. Any data collected by research teams must comply with Facebook’s terms and policies and have approval from the university’s institutional review board, if applicable. Please be aware that Facebook does not allow the automated scraping of public information from the platform.
Budget and payment
- Award amount can vary but is typically up to $60K.
- Payment will be made to the principal researcher’s host university as an unrestricted gift. Overhead is limited to 5 percent for gifts. Overhead amounts should be built into proposed budgets.
Timing and dates
- Applications are now closed.
- Notifications will be sent by email to selected applicants in April 2020.
Frequently Asked Questions
Do you typically limit the salary of the PI in the gift?+
Should the proposal be double- or single-spaced? Is there any required/expected font?+
What is the award cycle or when does the funding year begin and end?+
Can award funds be used to cover a researcher's summer salary while conducting research?+
Can you please explain the budget breakdown in more detail?+
We are working as co-PIs and are at the same institution. Is it possible to list both of our names as PI for an RFP proposal?+