Meta wants to use machine learning to promote fairness and anti-discrimination in advertising
In June 2022, Meta said it wanted to “expand its ad equity work.” On January 9, the company announced the launch of its Variance Reduction System (VRS, or gap reduction system) in the United States. machine learning, which should help spread ads more evenly across its apps. Initially, it will address real estate ads before expanding to job and loan ads.
Facebook has often been at the center of controversy and numerous complaints. The Quebec Court of Appeal has upheld a class action lawsuit brought on behalf of users who claimed Facebook discriminated against them because it allowed advertisers to target job and housing ads based on factors such as age, gender or race.
One such complaint, filed in August 2018 by the Assistant Secretary for Fair Housing and Equality with the U.S. Department of Housing and Urban Development (HUD), accused the company, now Meta, of violating the Fair Housing Act, an American federal law. In 1968, it banned discrimination in housing.
Therefore, the charge also relates to the targeting of real estate ads on the company’s platforms. Last June, Meta and HUD reached an agreement under which Meta specifically “ad targeting options that people may find sensitive”.
The company and the US Department of Justice worked together to develop the new VRS technology.
Meta said last June:
“Discrimination in housing, employment, and credit is a long-standing problem in the United States, and we are committed to expanding opportunities for marginalized communities in these spaces and elsewhere.”
He added:
“We’re making this change in part to respond to feedback we’ve heard from civil rights groups, policymakers and regulators about how our ad system serves certain categories of personalized ads, particularly those related to equity.”
VRS framework
SRV is a reinforcement learning framework that aims to minimize the gap between the number of people who see an ad and the larger target audience that might have seen it.
Reinforcement learning (reinforcement learning) is a form of ML that learns from trial and error to optimize toward a predetermined outcome — in this case, to minimize the difference in ad impressions across demographic subgroups, regardless of the reason for that difference.
Because of the settlement with the Department of Justice, advertisers cannot use targeting characteristics such as age, gender and zip code to determine the audience for their ads. SRV therefore receives aggregated data of demographic distributions and not this data for decision making.
Thus, when an ad is presented to a large enough panel, SRV measures the overall demographic distribution of people who see it and compares it to the demographic distribution of the appropriate target audience selected by the advertiser.
To respect individuals’ privacy, VRS relies on Bayesian Enhanced Surname Geocoding (BISG), a methodology developed by the RAND Corporation that combines surname data and geocoding to better estimate race and gender. ‘Ethnicity. It is built with privacy enhancements, including differential privacy, a technique that can help protect against the re-identification of individuals in aggregate data sets. According to Meta, VRS is a significant technological advance that shows how AI can be used responsibly to deliver personalized ads.
