top of page

Thanks for subscribing!

Search

Meta's Variance Reduction System: Is This the AI Fairness Solution We've Been Waiting For?

Writer's picture: Richard PaceRichard Pace

Updated: Nov 3, 2024


Meta's Variance Reduction System

Author's Note: For those more technically inclined, in June 2023 Meta researchers released a pre-print research paper "Towards Fairness in Personalized Ads Using Impression Variance Aware Reinforcement Learning" that provides further detail on the development and implementation of its Variance Reduction System.


* * *


In August 2022, I wrote about the US Department of Justice's ("DOJ's") landmark legal settlement with Meta Platforms ("Meta" aka Facebook) that resolved allegations of discriminatory advertising in violation of the Fair Housing Act. For readers unfamiliar with this "digital redlining" settlement and my views thereon, I refer you to my post "Algorithmic Bias and Alternative Data: Five Lessons From the DOJ's Meta Settlement" for relevant background.

A key open item from the DOJ's Settlement Agreement was a requirement for Meta to "de-bias" the machine learning algorithms within its digital ad delivery systems as - according to the DOJ's testing - such algorithms allegedly suppressed the display of housing-, employment-, and credit-related ("HEC") advertisements to certain protected class groups. The DOJ attributed such demographic disparities to "algorithmic bias" since Meta had already removed the algorithms' access to data attributes that directly or indirectly reflected a user's race/ethnicity and sex. Therefore, Meta's machine learning algorithms must still be inferring latent measures of user demographics from complex combinations of Meta's other user attributes.

In response to this DOJ requirement, Meta announced in January 2023 its implementation of a "Variance Reduction System" ("VRS") - a new artificial intelligence module for its HEC ad delivery system whose objective is to reduce the demographic disparities driven by this algorithmic bias. Along with this announcement, Meta released a blog post and technical paper providing additional details about the VRS's design and operation.

In its own simultaneous announcement related to Meta's Variance Reduction System, the DOJ stated,

"Federal monitoring of Meta should send a strong signal to other tech companies that they too will be held accountable for failing to address algorithmic discrimination that runs afoul of our civil rights laws.” and “This groundbreaking resolution sets a new standard for addressing discrimination through machine learning."

Given the DOJ's remarks about this "groundbreaking" system setting a "new standard" for addressing algorithmic discrimination, I explore in this blog post the key features of Meta's VRS in the context of AI-driven digital advertising, as well as what the VRS may imply more broadly about the government's views on algorithmic bias and potential remedies.


Let's dive in.


The Need For Meta's Variance Reduction System:

An Overview of Meta's Digital Advertising Process


Meta's Digital Ad Creation and User Targeting

Let's consider the case of an HEC advertiser who wishes to run a digital ad on Meta's social media properties. To start the process, the advertiser defines its advertising objective - such as driving users to its website or a landing page where they can apply for a job, loan, or residential lease. Next, the advertiser creates the digital ad content consisting of text, images, video, and other content whose objective is to engage the user and elicit the desired action. Finally, the advertiser defines its target audience - the pool of Meta's users who possess certain attributes that the advertiser values and that are typically related to customers considered the best match for the advertiser's offer. For HEC ads, Meta explicitly prohibits the selection of attributes that reflect directly or indirectly any prohibited basis characteristics such as race/ethnicity or sex.

As an example of this process, the HEC ad below started running on Facebook and Instagram in February 2023.

Digital ad for Credit Card
Source: Meta Ad Library, February 2023.

In this example, the ad's objective is to spur users to "Apply Now" for the bank's credit card product based on the ad's text, imagery, and offer. In the remainder of this post, I will generically refer to ads such as this as "the Bank's" ad as I intend to speak generally about such ads and do not wish to attribute my generalizations specifically to this MidFirst Bank ad of which I have no additional information (such as targeting criteria).

The Bank can target to whom this ad is displayed using attributes such as minimum age to legally contract (e.g., 18 or older), broad-based location (e.g., customers located within the Bank's market area)[1], as well as certain other non-demographic customer attributes collected by Meta on its users (e.g., people who are members of a personal finance Facebook group). With these targeting criteria, the Bank defines its "eligible audience" - that is, the potential pool of Meta users for whom this ad will be targeted. Keep in mind, however, that the eligible audience may be vast - containing millions or tens of millions of users - and the advertiser's budget may only allow it to display its ads to a fraction of such users. Also, there is a difference between the pool of potential eligible users and the pool of active eligible users that use Meta's platforms during the specific period of time over which the Bank's ad campaign is running. This distinction is important for measuring demographic variances within the VRS - as I will describe a bit later.


Meta's Digital Ad Delivery System

After completing the ad creation and user targeting process, the final step is Meta's delivery of the digital ad according to: (1) the advertiser's desired target audience (hereafter referred to as the "eligible audience" to be consistent with the DOJ's Complaint and Settlement Agreement), (2) overall ad budget for this campaign, and (3) desired ad campaign timing / length. I leverage the following graphic from Meta's VRS technical paper to describe how this part of the process works.


Meta's VRS
Source: "Toward fairness in personalized ads," Meta, 2023.

Once an ad campaign starts, when a member of the advertiser's eligible audience engages with Meta's platforms (e.g., Facebook or Instagram), ad space is inserted into that user's feed and - if using a web interface - other areas of screen real estate (see graphic below).


""

For each available ad space, Meta initiates an auction to display a specific ad to this user. The first step of this auction is to identify all potential ads that could be displayed - that is all active advertising campaigns for which this user represents a member of the eligible audience. In the Bank credit card ad example, the Bank's eligible audience likely has users that overlap with other advertisers' eligible audiences; therefore, the Bank must compete against those other advertisers to display its ad in this user's ad space at this time.

Let's say that there are 100 potential ads that could be displayed for this user in a given ad space. To determine which of the 100 potential ads will actually be displayed, Meta performs an ad auction that compares the Total Value of each advertiser's digital ad for this user. Again, leveraging graphics from Meta's technical paper, each potential ad's Total Value is calculated as follows.


Meta's VRS Ad Auction
Source: "Toward fairness in personalized ads," Meta, 2023.

The "$ bid" represents the amount the advertiser is willing to pay to display their ad in this ad space while the "estimated action rate" represents Meta's estimate of the likelihood that the user would actually take the advertiser's desired action if displayed their ad (i.e., the probability that the user would click on the "Apply Now" button within the Bank's ad). Since advertisers would prefer not to spend money displaying ads to users with a low likelihood of taking their desired actions, and because Meta would like to ensure that it is showing users "relevant" ads (i.e., those most likely to interest the user and elicit an action), Meta scales the advertiser's $ bid amount up or down by the estimated action rate.

For example, if the user is highly likely to engage with the ad, then the estimated action rate should be closer to 100% - thereby maximizing the advertiser's $ bid amount for this user's ad space. Alternatively, if the user is unlikely to be interested in this ad, then the estimated action rate would be closer to 0% - thereby lowering the advertiser's effective $ bid amount and reducing the likelihood that the advertiser will win this ad auction. According to Meta, a user's estimated action rate for a potential ad is generated by a set of machine learning models that "consider that person's behavior on and off Facebook (in accordance with the user's ad preferences and settings), as well as other factors, such as the content of the ad, the time of day, and interactions between people and ads."

The next component of an ad's Total Value is the Ad Quality metric. According to Meta, this represents "a determination of the overall quality of an ad, such as whether the ad is likely to be engagement bait or whether people are likely to provide negative feedback about the ad, such as repeatedly hiding or reporting it." An ad's quality metric is also estimated by machine learning models that "consider the feedback of people viewing or hiding the ad, as well as assessments of low-quality attributes (like too much text in the ad's image, sensationalized language, or engagement bait)."

Finally, Meta incorporates each advertiser's bidding strategy into the Total Value calculation (not shown in the graphic above) - that is, its desired ad pacing such as whether they desire their ad campaign to last for a certain time duration (e.g., four weeks) or whether they desire a campaign blitz due to the desire for quick responses and results (e.g., they are running a short-term sale or offer). In general, an advertiser's bidding strategy is implemented via a set of "pacing multipliers" that dynamically impact the advertiser's auction results by increasing or decreasing their bid values over time to match their bidding strategy.


For example, if the advertiser's bidding strategy is to extend its ad budget over a four-week campaign, then pacing multipliers may be implemented by Meta to slow the advertiser's delivered ad impressions (and corresponding ad spend) during early stages of the campaign. More specifically, these pacing multipliers would reduce the advertiser's bid amounts during periods where auctions may be particularly competitive (and, therefore, costly) so as not to exhaust the ad budget prematurely, and increase its bid amounts (where necessary and potentially up to a maximum cap) during subsequent "off-peak" periods.


Meta's VRS Pacing Multipliers
Source: "Toward fairness in personalized ads," Meta, 2023.

Combining all these components together - $ bid amounts, pacing multipliers, estimated action rates, and ad quality values - yields an ad's Total Value for a given user and a specific ad impression opportunity. The advertiser with the highest Total Value wins the auction and has their ad displayed. The entire process described above is fully automated and occurs in real time.

Meta's VRS: An Overview

What Is It?


In its June 2022 Complaint, the DOJ alleged that Meta's personalization algorithms - that is, the machine learning models underlying its ad delivery system (such as those producing the estimated action rates) - were creating illegal disparate impact in the display of HEC ads to Meta's users even without access to certain predictive attributes related to these users' prohibited demographic characteristics. More specifically, the DOJ alleged - with supporting evidence - that material variances existed between the sex and race/ethnicity distributions of users to whom certain HEC digital ads were targeted and those to whom its digital ads were actually delivered.

While an HEC advertiser cannot target their ads based on sex or race/ethnicity, there is still an underlying demographic distribution associated with its eligible audience. For example, suppose our Bank targets its credit card ad to users within the Bank's service area who are legally able to contract, and who are members of a personal finance Facebook group. While sex is not a targeting criteria, we can discern - after the eligible audience is created - that such users tend to be, on average, 55% male and 45% female.[2] Now, once the Bank's digital ad campaign is launched, it is possible that - even though the ads are delivered to users within the Bank's designated target audience - the actual sex distribution of the eligible user subset to whom the ads are displayed is materially different than the population of users in the Bank's eligible audience.[3]

While such demographic variances can occur for benign reasons, the DOJ's concern is that Meta's machine learning algorithms may still contain some latent bias even without access to user attributes that may serve as demographic proxies. While not explicitly stated in the DOJ or Meta settlement documents, such biases may lead to - for example - underestimated action rates for female members of the Bank's eligible audience (relative to male members) - thereby lowering the Total Values for HEC ads to female users and reducing the Bank's likelihood of winning their ad auctions. In the presence of such biases, we may observe potentially material variances in the sex distribution of the Bank's displayed ads versus that of the eligible audience (e.g., displayed ads could be 70% male and 30% female versus 55% male and 45% female in the eligible audience).[4]

To address these allegations of latent algorithmic bias, Meta developed the VRS - an AI-based module for its HEC ad delivery system that interacts with Meta's automated ad auction process to: (1) monitor the underlying demographics of an ad campaign's delivered ad impressions, (2) identify variances between these delivered ad demographics and the expected demographics of the ad campaign's active eligible users, and (3) implement corrective action strategies over the course of the ad campaign to reduce the demographic disparities below an agreed-upon level with the DOJ.

How Does It Work?


While Meta's technical paper provides a fair amount of descriptive information about its VRS, there are some details not discussed that prevent a comprehensive understanding of its operations. Nevertheless, given existing information, the following represents my current understanding of how it achieves its goal of reducing demographic disparities in HEC ad delivery.


VRS Training


Meta's VRS is trained on historical HEC ad campaign data, for which the underlying AI algorithm is tasked with finding an optimal set of "boosting multipliers" (i.e., adjustments to an ad campaign's pacing multipliers over its remaining term) to reduce measured variances between the demographic distributions of users to whom the ad was actually displayed versus the set of eligible users to whom the ad could have potentially been displayed. The VRS calibrates these boosting multipliers by:

  1. Calculating the ad's existing demographic variances at certain points in the ad campaign timeline.

  2. Trying different sets of boosting multiplier values for the remaining term of the ad campaign to increase the likelihood of winning auctions for certain eligible users.

  3. Periodically evaluating the VRS's progress in reducing demographic variances after sufficient ad impressions have been delivered.

  4. Updating the boosting multiplier strategy in response to (3), and

  5. Calculating the final demographic variances at the end of the campaign term.

During training - and conditional on starting variance levels and the remaining campaign term, the AI algorithm learns how to leverage the non-demographic user features, among other data, to derive optimal boosting strategies that reduce demographic variances below certain levels. Importantly, during this training process the VRS has no direct knowledge of users' individual demographics and, therefore, cannot derive the boosting multipliers based on these demographics. Instead, the VRS has access to certain summarized features of the eligible users entering the ad auction process - features that are derived from the subset of the user-ad interaction features driving the machine learning models for the "estimated action rate". Based on a "smart" trial and error process, the AI algorithm learns how to vary the boosting multipliers based on these user-ad features so as to reduce the campaign's overall demographic variances.

VRS Implementation

Once trained, Meta is able to deploy the VRS to reduce demographic differences in HEC ad delivery. For a given ad campaign, as ad impressions are delivered to eligible users, the VRS periodically intervenes to measure the real-time demographic variances between the users to whom the ad was delivered so far and the eligible users that could have been selected to view the ad during that period of time. Based on these measured variances, the VRS will adopt a specific boosting strategy on future ad auctions in an attempt to reduce these demographic variances in a demographically-blind manner (i.e., without knowledge of future eligible users' demographic information). That is, for each future ad auction, the VRS will choose whether to boost the pacing multiplier (thereby increasing the likelihood that the advertiser wins the auction for that user) based on the eligible user's user-ad interaction features and the learned relationship of these feature values during AI training with subsequent demographic variance reduction.

Essentially, the VRS appears to act as a separate algorithmic-based de-biaser of the machine learning models underlying the estimated action rates. That is, since those machine learning models are the likely source of any measured demographic variances in HEC ad delivery, the VRS uses a subset of key data inputs from those models to create a separate algorithm that effectively reverses the effects of this bias during the individual ad auctions.

Once the VRS implements a set of boosting multipliers, a certain time period is allowed to elapse to accumulate ad auction results that have been impacted by these adjustments. Then, the VRS intervenes again to update its measurement of accumulated demographic variances. Based on these new variance levels, and the remaining ad campaign term, the VRS updates its boosting strategy - again, with the goal of achieving final campaign demographic variances below an agreed-upon level. This process of VRS intervention and strategy revision occurs multiple times during the course of an ad campaign.

My Take: Three Key Observations


Meta's VRS certainly represents a technical contribution to the fight against algorithmic bias and its solution, being part of a formal DOJ legal settlement, has the added benefit of an imprimatur from one of the leading U.S. consumer protection enforcement agencies. So, is this the AI fairness solution we've been waiting for? Below I share some key observations to consider.

""

Key Observation: The DOJ did not object to Meta's use of aggregated demographic data to de-bias its consumer-facing machine learning algorithms.


As described previously, the VRS's goal is to minimize demographic distribution disparities between an ad's active eligible audience and the specific audience to whom the ad is actually delivered. To achieve this goal, the VRS utilizes aggregated demographic information on Meta's users to measure current disparities, to select appropriate boosting strategies to course-correct ad delivery - conditional on these disparities, and to measure whether its actions have had the desired fairness effect on ad campaign outcomes. Importantly, no individual-level demographic data is used during VRS training or implementation.

The DOJ's non-objection to Meta's use of aggregated demographic data to remove algorithmic bias is important as it may provide a roadmap to address evidence of algorithmic bias in other consumer-facing AI applications - such as credit scoring. As discussed in more detail in my previous blog post "Fools Gold? Assessing the Case For Algorithmic De-Biasing", current popular algorithmic de-biasing methodologies used in financial services - such as adversarial de-biasing and fairness regularization - work by leveraging demographic data to encode a latent form of "reverse disparate impact" into the credit scoring model's weights to yield more demographically-balanced model outcomes (e.g., loan approvals).

However, up until the DOJ's Meta settlement, the permissibility of such demographic data usage was an open question - with significant concerns among consumer lenders that such usage potentially violated applicable federal fair lending laws and regulations. Now, with the DOJ's Meta Settlement and its non-objection to the VRS's use of aggregate demographic data, consumer lenders may have a precedent for a regulatory-compliant use of such data to address disparate impact concerns. Of course, it's possible that the DOJ's non-objection is based - in part - on relevant differences between the Fair Housing Act (on which Meta's settlement is based) and the Equal Credit Opportunity Act, and/or the fact that the algorithm leveraging the demographic data is developed separately from the algorithm containing the alleged algorithmic bias - providing both separation and transparency in its usage. Unfortunately, the basis for the DOJ's apparent comfort is currently unknown and - while this is a promising development - lenders should still consult legal counsel before taking any actions in which any form of demographic data is used in the development of consumer credit models.


""

Key Observation: The DOJ did not object to Meta's "shuffle distance" method to measure disparities in multi-category demographic distributions - such as race/ethnicity - nor Meta's proposed maximum disparity thresholds for residual algorithmic bias.


In the Meta matter, the DOJ expressed concern over observed disparities in the demographic distribution of delivered ads relative to the expected demographic distribution of the advertiser's target audience. For binary demographic identities, such as sex, the measurement of such distributional disparities is straightforward - if the target audience is 45% female / 55% male and the audience to which the ad was delivered was 40% female / 60% male, then the resulting sex disparity is 5%. However, when the demographic identity is multi-category - such as for race/ethnicity, the overall distributional disparity is more complex. For example, consider the two race/ethnicity distributions displayed in Table 1 below for an advertiser's target audience and the final delivered audience (i.e., the users to whom the ad was actually shown).


""
Table 1

While one can calculate disparities for individual race/ethnicity categories - for example, -3% for Black or -5% for Hispanic - how does one calculate an overall aggregate disparity level across all categories of the distribution? That is, how far "off" is the demographic distribution of the delivered audience relative to the target audience? This question is particularly relevant when - unlike Meta's data on user sex - user race/ethnicity is unknown and is therefore estimated in the aggregate using the BISG proxy methodology.[5]

Based on these considerations, Meta proposed for the DOJ settlement to use "shuffle distance" to measure aggregate race/ethnicity distributional disparity levels which they define as "... the minimum fraction that needs to be moved (or shuffled, hence the name) from an actual distribution ... to match a reference distribution." In Table 1 above, the delivered audience would correspond to the "actual distribution" and the target audience would correspond to the "reference distribution".

To illustrate the calculation of shuffle distance for the two distributions in Table 1, assume the two demographic distributions are based on 100 ad impressions. The third column of Table 2 below then shows the corresponding build-up of the shuffle distance across all four race/ethnicity categories. For example, the delivered audience contained an estimated 7 Black users (= 7% x 100) - which is 3 fewer than the expected 10 Black users (= 10% x 100) based on the target audience demographics. Comparing these user variances across all four race/ethnicity categories, we see that the overall distributional disparity could be eliminated by "shuffling" 8 ad impressions across the four race/ethnicity groups - specifically, by taking 8 ad impressions away from White and All Other users (i.e., 6 from White and 2 from All Others) and reallocating them to Black and Hispanic users (i.e., 3 to Black and 5 to Hispanic). Accordingly, the minimum amount of ad impression "shuffling" needed to cure the overall distributional disparity is equal to 8% - equal to one half of the sum of the absolute values of the individual variances (= 16/2 in this example) divided by the total sample size (i.e., 100).


Meta VRS Shuffle Distance
Table 2

Meta and the DOJ pair this demographic distributional disparity metric with agreed-upon compliance thresholds to resolve the HEC ad delivery system's algorithmic bias concerns. Specifically, according to the DOJ's January 2023 press release:

... by Dec. 31, for the vast majority of housing advertisements on Meta platforms, Meta will reduce variances to less than or equal to 10% for 91.7% of those advertisements for sex and less than or equal to 10% for 81.0% of those advertisements for estimated race/ethnicity.

This indicates that the DOJ is willing to tolerate a maximum residual 10% distributional disparity for both sex and race/ethnicity (i.e., after the application of Meta's VRS). However, it also currently permits limited exceptions to this maximum threshold - permitting a maximum 8.3% (19.0%) of HEC ad campaigns to exceed the 10% maximum disparity level when evaluating compliance with expected sex (race/ethnicity) distributions. Such exceptions are likely driven by campaigns with small ad budgets and, therefore, (1) fewer opportunities for the VRS to course-correct measured demographic disparities, and (2) small numbers of ad impressions where even small "shuffling" of ad impressions can cause relatively large swings in distribution values.

In general, while the specific distributional disparity metrics (i.e., shuffle distance) and compliance thresholds (e.g., maximum 10% residual disparity) contained in this settlement pertain to the facts and circumstances of the Meta Complaint and the alleged Fair Housing Act violations, they are also helpful data points to consider when evaluating broadly similar algorithmic bias risks in other consumer-facing areas - such as consumer lending.

While the previous two key observations yield helpful information related to: (1) measuring algorithmic bias (i.e., shuffle distance), (2) demonstrating how aggregate demographic data may be used non-objectionably to de-bias algorithms, and (3) indicating potential "practical" significance thresholds for residual bias levels (i.e., 10%), there is also an important potential pitfall with the VRS de-biasing approach - to which I now turn.


""

Key Observation: Meta's AI fairness solution effectively eliminates any demographic differences in users' interests in specific HEC ads - thereby potentially: (1) displaying ads to users for whom they are not relevant, (2) displacing other ads to such users that would be relevant, and (3) reducing advertisers' ROIs by generating fewer ad conversions.


As discussed previously, Meta's VRS uses the advertiser's target audience demographics as the baseline to which to measure potential bias in delivered ads. However, these demographics abstract from other considerations that determine whether a user is actually interested in the ad. That is, the subset of eligible users who are actually interested in an ad may differ materially in attributes from the larger group of all eligible users. For example, people who tend to respond to credit offers may more likely need credit and, therefore, may be of lower income or wealth - thereby skewing demographically by age and/or race/ethnicity due to the underlying correlation. People who respond to certain employment ads - such as Registered Nurses - will need to have that occupational qualification and such individuals may be heavily skewed demographically by sex. Such interest- or qualification-driven demographic variances are effectively ignored by the VRS - even if legally permissible[6] - as it focuses instead on the more diffuse aggregate-level eligible audience demographics.

However, if we were to consider the relevancy of an ad to an advertiser's eligible users (which is a primary benefit of digital advertising), we see that the baseline demographic distribution to which the VRS is adjusting is not necessarily the demographic distribution of users who are actually interested in (or qualified for) the ad's subject. For example, the eligible audience for the Registered Nurse employment ad is likely much more balanced by sex than the subset of this audience that actually possesses the occupational qualification for the job.[7] However, Meta's VRS does not permit demographic differences based on HEC ad relevancy. It doesn't matter that higher income / higher wealth users are much less interested in the Bank's credit card offer if such ad irrelevancy is correlated with user demographics. Similarly, it doesn't matter that users without certain occupational qualifications are much less interested in employment ads for those positions if the absence of such qualifications is correlated with user demographics. In fact, any demographic variance from the eligible audience that is created by the estimated action rate models (i.e., the models that are designed to measure ad relevancy) is considered part of the "algorithmic bias" that must be removed.[8]

What this means is that the VRS may actually disadvantage some users (even those in protected class groups) by showing them ads in which they are not interested, or for which they are not qualified. And this comes at a cost since, instead of this ad, these users could have been shown alternative ads that were more relevant to them. This calls into question whether users (including protected class users) actually benefit, on net, from the VRS. This is because we do not know what portion of the "shuffle distance" is due to true algorithmic bias, and which portion is due to legitimate differences in user interest or qualifications. While those users in the former group benefit from Meta's VRS, those in the latter group may not - and we do not know how these two groups net out.

Beyond the user impact, I also note that an advertiser's cost per acquisition may now be higher than before because it is now showing its ads to a greater number of individuals who may not find them relevant. As a result, the return on its ad spend may be much lower then before - thereby causing some advertisers to switch away from Meta and toward other delivery channels with greater advertising efficiency. It would be very useful if Meta shared the impact of VRS on advertising efficiency metrics to see how important this ad relevancy issue really is.


Final Thoughts

From a pure digital marketing perspective, Meta's VRS provides a means to address algorithmic bias in HEC ad delivery that satisfies the DOJ's requirements. However, it achieves this goal with an important trade-off - which is a potentially material reduction in digital advertising efficiency.[9]

Taking a broader view, Meta's VRS suggests that the use of aggregate demographic data to address algorithmic bias in other consumer-facing models (such as credit models) may not be as verboten as currently considered - giving a glimmer of hope to the potential permissibility of popular credit model de-biasing methodologies such as adversarial de-biasing and fairness regularization. However, there may be important differences between the Fair Housing Act, which underlies the DOJ's Meta Settlement, and the Equal Credit Opportunity Act that may explain the DOJ's unexpected position in this case. Clearly, more guidance and legal analysis is needed here. Additionally, even if the use of aggregate demographic data to de-bias credit models is deemed permissible, consumer lenders need to ensure that such de-biasing does not create other unintended regulatory risks.

In terms of whether Meta's VRS algorithm is the broader-based AI fairness solution we have been looking for, that's not entirely clear at the moment. While it provides some interesting food for thought regarding the use of demographic data for model de-biasing, and potential practical significance thresholds for residual demographic variances, its approach does not necessarily fit with a broad range of consumer-facing use cases. This is because it is designed to minimize the variances between an actual demographic distribution and a reference or expected demographic distribution - for which the number of use cases may be limited. While this structure may fit digital advertising and - more broadly - marketing response models, it is not clear that there is a direct applicability to credit models - although time will surely tell.

* * *


ENDNOTES:


[1] For HEC ads, any location targeting must have a minimum 15-mile radius. More broadly, Meta addressed a number of alleged Fair Housing Act violations related to its HEC target audience creation process in both a 2019 settlement with the National Fair Housing Alliance as well as in the current DOJ settlement.


[2] There is also an underlying race/ethnicity distribution - along with a corresponding risk of material variances according to this demographic characteristic as well. However, for the purposes of this example, I focus solely on sex. I should also point out that the VRS only focuses on potential disparate impact in ad delivery; importantly, it is not designed to address potential disparate impact in the demographic distribution of the eligible audience (Meta's other legal settlement requirements addressed that concern).


[3] In the extreme, where an advertiser could reach 100% of its eligible audience, this disparate impact concern would be moot since the two demographic distributions would be the same. However, because (1) an advertiser's limited budget permits ad delivery for only a subset of its eligible audience, (2) that subset is not a random sample of eligible users, and (3) all eligible users may not be active during the ad campaign window, the potential exists for the two demographic distributions to differ - potentially materially.


[4] Technically, the sex distribution of active users who are eligible to see the Bank's ad may differ slightly from the original target sex distribution due to fluctuations in platform usage, spikes in platform activity tied to external events, and other factors. Because of this, Meta's VRS measures variances relative to the sex distribution of active eligible users during the specific campaign period - not the sex distribution associated with the theoretical target audience based on Meta's overall population of users.


[5] According to Meta's Technical Paper, it uses a privacy-enhanced version of the Bayesian Improved Surname Geocoding ("BISG") proxy methodology to estimate race/ethnicity. See, in particular, "How Meta is working to assess fairness in relation to race in the U.S. across its products and systems" Meta Technical Report, November 2021. Additionally, according to Guidehouse's VRS Compliance Metrics Verification Report dated June 29, 2023, Meta uses a 50% BISG probability threshold to assign users to a specific race/ethnicity category.


[6] I am not taking a position on the legality of these qualification- or interest-based demographic differences as that is the domain of legal counsel. However, to the extent that these demographic differences are legally permissible, they are not considered as such by Meta's VRS.


[7] For example, the U.S. Census reports that women comprise about 85% of Registered Nurse positions.


[8] Importantly, Meta is not acknowledging that its estimated action rate models are biased because they systematically over- or under-predict actual action rates by demographic group. Instead, they acknowledge that the model's estimated action rates vary across demographic groups - even if such variations are consistent with underlying user behavior and are not subject to potential data input biases.


[9] I want to be clear that this reduction in advertising efficiency is not tied to the reduction in true algorithmic bias. It is, instead, driven by the VRS's goal of eliminating all demographically-correlated ad relevancy differences - some of which may be legitimate (as discussed). Of course, this assumes that these legitimate ad relevancy differences are also legally permissible under state and federal fairness laws and regulations.


© Pace Analytics Consulting LLC, 2023.

616 views
Share your feedback on the AI LendScape Blog
Please rate your overall satisfaction with our blog content
Very dissatisfiedA bit dissatisfiedPretty satisfiedSatisfiedVery satisfied

Thanks for sharing!

Your feedback is anonymous.

© 2025 by Pace Analytics Consulting LLC

The information presented herein does not constitute financial or other professional advice and is intended to be general in nature. It does not take into account your specific circumstances and should not be acted on without a full understanding of your current situation and future goals and objectives by a fully qualified advisor.

bottom of page