Equalizing Credit Opportunity in Algorithms Aligning Algorithmic Fairness Research with U.S. Fair Lending Regulation_2

2025-04-29 1 0 548.11KB 12 页 10玖币

侵权投诉

Equalizing Credit Opportunity in Algorithms: Aligning

Algorithmic Fairness Research with U.S. Fair Lending Regulation

I. Elizabeth Kumar

Brown University

USA

iekumar@brown.edu

Keegan E. Hines

Arthur AI

USA

keegan@arthur.ai

John P. Dickerson

Arthur AI

USA

john@arthur.ai

ABSTRACT

Credit is an essential component of nancial wellbeing in America,

and unequal access to it is a large factor in the economic dispari-

ties between demographic groups that exist today. Today, machine

learning algorithms, sometimes trained on alternative data, are in-

creasingly being used to determine access to credit, yet research has

shown that machine learning can encode many dierent versions

of “unfairness,” thus raising the concern that banks and other nan-

cial institutions could—potentially unwittingly—engage in illegal

discrimination through the use of this technology. In the US, there

are laws in place to make sure discrimination does not happen

in lending and agencies charged with enforcing them. However,

conversations around fair credit models in computer science and

in policy are often misaligned: fair machine learning research of-

ten lacks legal and practical considerations specic to existing fair

lending policy, and regulators have yet to issue new guidance on

how, if at all, credit risk models should be utilizing practices and

techniques from the research community. This paper aims to better

align these sides of the conversation. We describe the current state

of credit discrimination regulation in the United States, contextu-

alize results from fair ML research to identify the specic fairness

concerns raised by the use of machine learning in lending, and

discuss regulatory opportunities to address these concerns.

CCS CONCEPTS

•Computing methodologies →Machine learning

;

•Applied

computing →Law, social and behavioral sciences.

ACM Reference Format:

I. Elizabeth Kumar, Keegan E. Hines, and John P. Dickerson. 2022. Equalizing

Credit Opportunity in Algorithms: Aligning Algorithmic Fairness Research

with U.S. Fair Lending Regulation. In Proceedings of the 2022 AAAI/ACM

Conference on AI, Ethics, and Society (AIES’22), August 1–3, 2022, Oxford,

United Kingdom. ACM, New York, NY, USA, 12 pages. https://doi.org/10.

1145/3514094.3534154

INTRODUCTION

Credit is an essential component of nancial well-being for Ameri-

cans, and unequal access to it is a signicant factor in the economic

Permission to make digital or hard copies of all or part of this work for personal or

classroom use is granted without fee provided that copies are not made or distributed

for prot or commercial advantage and that copies bear this notice and the full citation

on the rst page. Copyrights for components of this work owned by others than the

author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or

republish, to post on servers or to redistribute to lists, requires prior specic permission

and/or a fee. Request permissions from permissions@acm.org.

AIES’22, August 1–3, 2022, Oxford, United Kingdom

ACM ISBN 978-1-4503-9247-1/22/08. . . $15.00

https://doi.org/10.1145/3514094.3534154

disparities between demographic groups that exist today. For this

reason, it is critical to make sure the American lending ecosys-

tem is free of discrimination. In America, there are laws in place

which specically ban discrimination in lending, as well as agen-

cies charged with enforcing them. Today, machine learning (ML)

algorithms (sometimes trained on “nontraditional” data) are increas-

ingly being used to allocate access to credit. A vast body of research

has demonstrated that ML algorithms can encode many dierent

versions of “unfairness,” thus raising the concern that banks and

other nancial institutions could—potentially unwittingly–engage

in illegal discrimination through the use of this technology.

The nebulous threat of “algorithmic discrimination” poses a

challenge to federal regulators, who must decide how, if at all, to

update their enforcement practices or issue new guidance in light

of these concerns [

], which are often articulated by computer

scientists in the abstract and not in terms of the actual practices,

data, and algorithms used in this sector. Meanwhile, without specic

guidance from regulators, researchers and practitioners who want

to study or apply fair ML in this particular setting lack a clear

picture of the kinds of tools and metrics that will be useful, legal,

and practical for detecting and correcting unfairness in algorithms

in this setting. For these reasons, this paper aims to orient the

conversation around fair ML research in the context of predicting

credit risk from both perspectives.

In Section 1, we briey describe the state of American fair lend-

ing regulation and analyze recent messaging from certain federal

agencies on the threat of algorithmic discrimination. In Section 2,

we discuss methods proposed by the ML community to measure un-

fairness in algorithms, and determine the extent to which they may

relate to the principles of the Equal Credit Opportunity Act (ECOA)

and the goals of the federal agencies discussed above. Keeping these

metrics in mind, we contextualize results from fair ML research

in the consumer credit setting, and identify specic fair lending

risks throughout dierent parts of a machine learning system’s

development. By analyzing how these mechanisms are likely to

play out in the credit setting, we can be more specic about the

kinds of problems regulators should anticipate and address, rather

than repeating the folk wisdom of "bias in, bias out." Finally, in

Section 3, we discuss specic opportunities for regulators to use

their authority to encourage fair ML practices.

1 CREDIT DISCRIMINATION REGULATION

IN THE UNITED STATES

In this section, we provide a background on the laws and policies

which regulate anti-discrimination in consumer credit. We further

set the stage for the conversation about algorithmic discrimination

by identifying specic comments and actions from federal agencies

arXiv:2210.02516v1 [cs.LG] 5 Oct 2022

signifying their willingness to tackle the issue of discrimination in

algorithms.

1.1 Fair lending legislation

1.1.1 Anti-discrimination legislation. The issue of discrimination

in credit lending decisions is not novel to the algorithmic setting.

While lending has been around for centuries, Americans increas-

ingly began to rely on consumer credit to nance large purchases

in the 1950’s and 60’s [

]. During this period, individual loan o-

cers and specialists were ultimately responsible for the subjective

determination of whether a loan applicant was creditworthy; nu-

merical methods for estimating credit risk existed but were not

widely or systematically used [

]. This presented a risk of in-

tentional discrimination due to personal bias. Additionally, some

codied lending policies in eect at the time clearly disadvantaged

women and minorities. During congressional hearings, testimonies

cited practices such as requiring single women to provide a male

co-signer for a mortgage loan [40, 81].

In the spirit of implementing ideas from the civil rights legislation

of the 60’s, which did not directly address lending, ECOA was

passed in 1974 to ensure that all Americans were treated fairly in

a system that determined so much of their economic success. It

prohibits creditors from “discriminat[ing] against any applicant,

with respect to any aspect of a credit transaction on the basis of

race, color, religion, national origin, sex or marital status,” among

other factors [

]. The law applies to any organization that extends

credit, including loans and credit cards.

The Fair Housing Act, also known as Title VIII of the Civil Rights

Act of 1968, prohibits discrimination in housing on the basis of sev-

eral protected characteristics, and applies to mortgage providers.

The U.S. Department of Housing and Urban Development (HUD) en-

forces the Fair Housing Act, and has specied narrow rules making

disparate impact litigation dicult; partly because of this, mortgage

algorithms are not our main focus in this paper.

1.1.2 Data collection rules. At the time of its passing, the ECOA

gave the Board of Governors of the Federal Reserve Board (FRB)

rulemaking authority to implement the law; this set of rules is

known as Regulation B. Regulation B specically prohibits the

collection of information about protected characteristics: "A creditor

shall not inquire about the race, color, religion, national origin, or

sex of an applicant or any other person in connection with a credit

transaction" [

]. Credit transactions, here, can include things like

consumer credit, business credit, mortgage loans, and renancing.

A glaring set of exceptions to this rule are in cases where the

Home Mortgage Disclosure Act (HMDA) applies. Passed in 1975, the

act requires certain nancial institutions to provide mortgage data

to the public, and in particular requires lenders to collect and report

race and gender information of mortgage applications. The act was

drafted in response to the practice of redlining, in which lenders

would explicitly identify geographic regions and neighborhoods

that they would not lend to because they were inhabited by people

of color. This information is used to identify indicators of mortgage

discrimination and encourage lenders to comply with ECOA [84].

In the non-mortgage setting, Regulation B contains an additional

exception to the ban on collecting protected characteristics: when

the information is explicitly collected for self-testing, which is

dened as any inquiry “designed and used specically to determine

the extent or eectiveness of a creditor’s compliance with the Act”

[

]. In doing so, lenders must make it clear to the applicant that

providing the information is voluntary. However, this practice is

very uncommon; Slaughter et al

. [83]

speculate that this is because

of a "fear that their collection of the data will validate or exacerbate

claims that their decisions are biased." Self-testing might also be

disincentivized if corporations believe that the data itself would

ultimately benet plaintis in a potential disparate impact suit.

It may seem counter-intuitive that HMDA requires the collection

of sensitive information while ECOA bans it. In fact, both HMDA’s

requirement for collecting sensitive information and ECOA’s ban

on it are controversial. Some argue that the existence of HMDA

provides an important basis of evidence for lawsuits or that the pol-

icy itself caused lenders to curb their own discriminatory practices,

and thus a similar provision should be in place for non-mortgage

lenders [

]. Others, especially banks, have argued that

HMDA is unfair, costly, and leads to false accusations of illegal

discrimination [54].

The Federal Reserve Board, which was responsible for enforcing

ECOA until the Consumer Financial Protection Bureau (CFPB) was

established, has considered removing the ban on the collection of

protected information several times since the law was originally

passed. In 2003, it ultimately rejected a proposal to lift the ban

and mandate the collection of certain sensitive information [

The rst reason it cited was the natural one: that creditors might

use this information for discriminatory purposes; however, many

members of Congress, consumer advocates and researchers found

this unconvincing [

]. The second was that "many creditors would

elect not to collect the data while those that did collect it would

use inconsistent standards, criteria and methods. Consequently, the

data would be of questionable utility because there would be no

assurance of its accuracy nor would there be any way to compare it

from creditor to creditor" [

]. The U.S. Government Accountability

Oce found in 2008 that while such a mandatory data collection

could provide benets to researchers and regulators, it could be

costly or dicult for the lenders themselves [

]; Bogen et al

. [14]

suggest that the failure to implement such measures has largely

been due to pressure from banks, which Taylor

[84]

found were the

overwhelming dissenting voice in responses to the FRB’s request

for comments on their proposal.

1.2 Fair lending in practice

The two major discrimination doctrines which are relevant to fair

lending law today are disparate treatment and disparate impact.

Disparate treatment applies when individuals are explicitly treated

dierently on a prohibited basis. Under disparate impact doctrine,

on the other hand, a creditor may be found to have illegally dis-

criminated against a protected class if the eect of the practice

adversely impacts that group even if the policy in question was

facially neutral. The Supreme Court has found that the disparate

impact is cognizable under the FHA [

], but has not made a similar

ruling about ECOA. However, the court’s language in Inclusive

Communities [

], relevant case law [

], and the CFPB’s ocial

interpretation of Regulation B [

] all support the general consen-

sus that disparate impact theory is cognizable under ECOA. Federal

courts have consistently upheld this since the 1980s [23].

Plaintis usually rely on burden-shifting systems for establishing

aprima facie claim under both theories, which can then be rebut-

ted by the defendant. For a disparate treatment case, most circuit

courts have found that a modication of the McDonnell-Douglas

test, originally developed by the Supreme Court in an employment

discrimination case [

], can be applied to an ECOA claim–but there

is no ocial nationwide rule on the issue [

]. In a prima facie

disparate impact claim [

], a plainti must point to a specic policy

or action taken by the defendant that had a disproportionately ad-

verse impact on members of a protected class. The defendant may

respond by arguing there is a legitimate business necessity for the

policy. Then, the plainti can respond by arguing there was a less

discriminatory alternative that the defendant refused to use.

In a disparate impact claim, expert statistical testimony is neces-

sary to demonstrate that an adverse impact exists and is dispropor-

tionately felt by members of a protected class [

]. Again, we lack

ocial Supreme Court guidance on how exactly to go about this

under ECOA. In employment discrimination cases, however, the

ratio of the proportion of protected class that receives a favorable

outcome and the proportion of the control class is used; the oft-

cited "80% rule" is related to this statistic [

]. A related metric with

precedence in the credit setting is the standardized mean dierence

of outcomes for two groups [

]. However, it is usually insucient

to simply compare the approval rates of two groups of applicants;

since information related to creditworthiness is generally available,

higher courts generally expect that plaintis will compare the se-

lection rates of qualied applicants [

]. For this reason, statistical

evidence which controls for drivers of creditworthiness–such as a

conditional marginal eects test–are seen as more appropriate by

federal agencies [19].

It is dicult for plaintis to nd evidence that an individual lend-

ing decision was discriminatory, especially in the non-mortgage

setting where sensitive attribute data about applicants is generally

unavailable; some lower courts have historically acknowledged this

[

]. Bogen et al

. [14]

point out that "one of the few, robust public

studies on credit scores and discrimination in the United States

was performed by the FRB in 2007, at the direction of Congress.

To conduct its analysis, the FRB created a database that, for the

rst time, combined sensitive attribute data collected by the Social

Security Administration with a large, nationally representative sam-

ple of individuals’ credit records... this unusual undertaking would

not have been possible without signicant governmental time and

resources." Interestingly, the CFPB has worked around this data

limitation in some of its enforcement actions by imputing racial

information using Bayesian Improved Surname Geocoding (BISG)

to amass evidence of disparate impact [

]. On the other hand, in the

mortgage setting where data is available, HMDA data alone cannot

prove or disprove discrimination, and the results of discrimination

studies using HMDA data are usually contentious [75].

An important precedent is, of course, the general acceptance of

traditional credit scores as a basis of loan underwriting. Like the ma-

chine learning algorithms which are the focus of this paper, credit

scores are functions of data which are meant to provide a quantita-

tive basis on which to make a lending decision. As of yet, there have

not been successful challenges against credit scores using disparate

impact theory [

]. A combination of factors has contributed to this,

but one seems to be that ocial CFPB interpretations of ECOA and

OCC guidance on models are fairly generous as to what counts as

a business necessity and relation to creditworthiness, respectively

[

]. Further complicating this matter is the fact that creditors tend

to (credibly) argue that their scoring methods tend to expand credit

to minority applicants when compared to other methods. The FRB

bolstered the credit score’s ubiquity in their analysis of the 2007

database: they claimed that while credit scores have a “dierential

eect" [

], they did not "produce a disparate impact" [

] because

credit characteristics do not act as "proxies for race and ethnic-

ity" according to their own denition (which we will discuss the

limitations of in Section 2).

1.3 Agency communications on fair lending in

algorithms

In this section, we analyze recent messaging from several federal

agencies on the threat of algorithmic fairness in nance and credit.

These agencies are generally allowed to operate independently, but

many have been known to act cooperatively and take a unied

stance on the interpretation of the law [

]. The OCC, FRB, FDIC,

and CFPB recently issued a rare joint request for information re-

garding the use of articial intelligence (AI) in nancial services,

inquiring, among other things, whether banks and other interested

parties feel that additional regulatory guidance on the matter is

necessary [

]. Their response to the threat of algorithmic discrim-

ination will be highly inuential since, as Alex Engler has argued,

"major legislative changes to AI oversight seem unlikely in the near

future, which means that regulatory interventions will set prece-

dent for the government’s approach to protecting citizens from AI

harms" [36].

1.3.1 The Consumer Financial Protection Bureau (CFPB). The CFPB

was created by the Dodd-Frank Wall Street Reform and Consumer

Protection Act in 2011. It was designed to consolidate responsibili-

ties from several other agencies such as the Federal Reserve, FTC,

and FDIC, to write and enforce rules for both bank and non-bank

nancial institutions. It has situated itself as being well-posed to

tackle new regulatory challenges introduced by technology. The

CFPB’s internal “Oce of Competition and Innovation," dedicated

in part to addressing these challenges, has taken initiatives such as

holding tech sprints, issuing no-action letters (NALs), and develop-

ing compliance assistance sandboxes.

The most relevant action the CFPB has taken with respect to

algorithmic discrimination was its NAL to ntech lending company

Upstart in 2017. Upstart provided detailed public (and some private)

information about its underwriting process with the bureau and

requested a formal statement from the CFPB that they would not

trigger any enforcement action [

]. The CFPB granted the NAL.

Part of the terms of the letter stipulated that Upstart would send

the CFPB updates "regarding the loan applications it receives, how

it decides which loans to approve, and how it will mitigate risk to

consumers, as well as information on how its model expands access

to credit for traditionally underserved populations" to "further its

understanding of how these types of practices impact access to

credit generally and for traditionally underserved populations, as

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

EqualizingCreditOpportunityinAlgorithms:AligningAlgorithmicFairnessResearchwithU.S.FairLendingRegulationI.ElizabethKumarBrownUniversityUSAiekumar@brown.eduKeeganE.HinesArthurAIUSAkeegan@arthur.aiJohnP.DickersonArthurAIUSAjohn@arthur.aiABSTRACTCreditisanessentialcomponentoffinancialwellbeinginAmerica...

展开>> 收起<<

Equalizing Credit Opportunity in Algorithms Aligning Algorithmic Fairness Research with U.S. Fair Lending Regulation_2.pdf

共12页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Equalizing Credit Opportunity in Algorithms Aligning Algorithmic Fairness Research with U.S. Fair Lending Regulation_2

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: