INTRODUCTION:
Meta had been working with the Department of Justice to develop a new technology to distribute ads without inequity on their apps like Facebook, Instagram, etc. And now it has launched the VRS in the US for housing ad and will expand it this year to employment and credit ads.
WHAT IS VARIANCE REDUCTION SYSTEM:
VRS is an offline reinforcement learning framework with the clear goal of minimizing the difference in number of ad views between people who have seen the word and the wider eligible audience who could have seen the ad.
WHAT IS THE PURPOSE OF VRS:
The main aim of VRS is to ensure that the audience who ends up seeing housing, employment, or credit ad more closely mirrors the eligible targeted audience for that ad. Meta does this by regularly measuring the actual audience for a particular ad to see how it compares with the demographic distribution such as age, gender, and estimated race of the audience the advertiser has chosen. The initial launch focuses on gender and estimated race. To respect people,s privacy, the system depends on privacy enhancing approaches.
HOW IT WORKS:
The demographic distribution of the eligible ratio of age, gender, and estimated racial distribution of the population of users targeted by an advertiser to view their ad is measured. As the ad is being casted, the demographic distribution of the impressions of the ad is measured periodically.
When there is a chance to show an ad to someone, all the ads that are eligible to be shown to that person are reduced by ad auction (calculating and comparing the total value of each ad). After that a process known as pacing adjusts the ad’s total value.
The VRS remeasures the audience’s demographic distribution and updates the pacing of ads throughout the campaign to reduce differences between the audiences. When there’s a new chance to show an ad to someone, the system uses the latest demographic measurements and the limited information about that person to adjust the pacing of the bid to encourage the ad to be delivered to the audience that more closely mirrors the ad’s eligible targeted audience.
PRIVACY APPROACHES:
By considering people’s privacy policy, Meta uses the following privacy- approaches while implementing VRS:
- VRS will not have access to the person’s age, gender, or estimated race.
- Estimated race will be measured using Meta’s privacy-enhanced implementation of Bayesian Improved Surname Geocoding.
Aggregate demographic measurements which are generated and used by VRS will involve distinctive privacy noise to prevent the system from learning and eventually acting on individual-level demographic information with high accuracy.