Simultaneous Improvement of ML Model Fairness and Performance by Identifying Bias in Data Aakash Agarwal Bhushan Chaudhari Dr. Tanmoy Bhowmik

2025-05-03 0 0 103.17KB 6 页 10玖币

侵权投诉

Simultaneous Improvement of ML Model Fairness

and Performance by Identifying Bias in Data

Aakash Agarwal, Bhushan Chaudhari, Dr. Tanmoy Bhowmik

AI Garage Mastercard

(aakash.agarwal,bhushan.chaudhari,tanmoy.bhowmik)@mastercard.com

Abstract

Machine learning models built on datasets containing discriminative instances

attributed to various underlying factors result in biased and unfair outcomes. It’s a

well founded and intuitive fact that existing bias mitigation strategies often sacriﬁce

accuracy in order to ensure fairness. But when AI engine’s prediction is used for

decision making which reﬂects on revenue or operational efﬁciency such as credit

risk modeling, it would be desirable by the business if accuracy can be somehow

reasonably preserved. This conﬂicting requirement of maintaining accuracy and

fairness in AI motivates our research. In this paper, we propose a fresh approach

for simultaneous improvement of fairness and accuracy of ML models within a

realistic paradigm. The essence of our work is a data preprocessing technique that

can detect instances ascribing a speciﬁc kind of bias that should be removed from

the dataset before training and we further show that such instance removal will have

no adverse impact on model accuracy. In particular, we claim that in the problem

settings where instances exist with similar feature but different labels caused by

variation in protected attributes, an inherent bias gets induced in the dataset, which

can be identiﬁed and mitigated through our novel scheme. Our experimental

evaluation on two open-source datasets demonstrates how the proposed method can

mitigate bias along with improving rather than degrading accuracy, while offering

certain set of control for end user.

1 Introduction

AI powered predictive modeling techniques have been widely adopted by business verticals in different

domains such as ﬁnance, healthcare, sports, banking, etc, often for making sensitive decisions ranging

from personalized marketing, loan application approval (Mukerjee et al. 2002) [3] to dating and

hiring process(Bogen 2018, Cohen 2019)[4,5]. Unsurprisingly, with the continuous evolution and

ever-increasing complexity, there have been several recent high-proﬁle examples of machine learning

(ML) going wrong in terms of bias, fairness and interpretability.

The presence of unintended demographic disparities or differential/disproportionate impact on

individuals by machine learning models is demonstrated by (Calders 2013) [29]. Unfairness can be

imparted in models because of bias present in training data. Various types of bias such as annotation

bias, historical bias, prejudice bias, etc may lead to unfair models and selective bias towards a

particular group. In (Mehrabi et al. 2019)[7], the authors have provided a comprehensive coverage of

such biases supportaed by real life examples.

The deﬁnitions used to understand the bias in models can be broadly categorized into three types:

independence, separation and sufﬁciency (Sharma et al.2020)[10]. Speciﬁcally, a classiﬁer satisﬁes

independence if the protected attribute (such as race or gender) for which the model may be biased is

independent of the classiﬁer decision. Separation is satisﬁed if the classiﬁer decision is independent

35th Conference on Neural Information Processing Systems (NeurIPS 2021), Sydney, Australia.

arXiv:2210.13182v1 [cs.LG] 24 Oct 2022

of the protected attribute conditioned on the true label. Sufﬁciency is satisﬁed if the true label is

independent of the protected attribute conditioned on the classiﬁer prediction. Details on these

fairness criteria, both mathematically and with respect to different worldviews, may be found in

(Barocas 2019, Yeom 2018) [27,28] along with deﬁnitions of fairness metrics, such as statistical

parity difference for independence and average odds difference for separation, from (Garg et al. 2020,

Bellamy et al. 2019, Sharma et al. 2020) [8,9,10]. We are taking statistical parity difference and

average odds difference metrics into consideration for this paper while being aware of the fact that

there are various fairness metrics which are relevant to gauge biasness of models. Determining the

right measure to be used must consider the proper legal, ethical, and social context.

Using these fairness metrics, several bias mitigation algorithms are developed to satisfy the various

criteria of fairness for machine learning models to reduce bias. Methods to mitigate bias generally

fall into three categories. Pre-processing techniques transform the data so that the underlying

discrimination is removed (Alessandro 2017)[21]. If the algorithm is allowed to modify the training

data, then pre-processing can be used (Bellamy et al. 2018)[22]. Our proposed methodology falls

under this category as explained in the subsequent sections. In-processing techniques try to modify

and change state-of-the-art learning algorithms in order to remove discrimination during the model

training process (Alessandro 2017) [21]. If the algorithm can only treat the learned model as a black

box without any ability to modify the training data or learning algorithm, then only post-processing

can be used in which the labels assigned by the black-box model initially get reassigned based on a

function during the post-processing phase (Bellamy et al. 2018)[22].

The literature extensively discusses the inherent trade-off between accuracy and fairness – as we

pursue a higher degree of fairness, we may compromise accuracy (see for example (Kleinberg et al.

2017) [12]). Many papers have empirically supported the existence of this trade-off (Be-chavod 2017,

Friedler et al. 2019) [13, 14]. Generally, the aspiration of a fairness-aware algorithm is to develop

a model that is fair without signiﬁcantly compromising the accuracy or other alternative notions of

utility.

In this paper, we are proposing a method to ﬁnd bias inducing samples in the dataset and then

dropping these samples such that the pre-processed dataset represents a more equitable world. In

an equitable world, model outcome is independent of the protected attributes (such as gender, race,

etc). Our novel approach can be described as follows: given a dataset that contains a protected

attribute (such as gender, race, etc), samples with similar attributes but different protected attributes

and different outcomes are ﬂagged. For example – Credit Risk dataset contains 2 samples: 1 male

and 1 female such that male and female sample have same attributes but model predicted low risk

for male and high risk for female. We establish that these instances induce bias as the attributes are

same but the outcome is different due to dependence on protected attributes (such as male, female)

and thereby result in unfair treatment by the model. Further, protected attributes are not used for

modelling, so such samples can confuse the model as they have nearly have the same attributes but

different label and thus can be viewed as pseudo label noise.

Hence our objective boils down to detect and remove such instances before training to make sure

the resultant model is fairer. In the process we also show that such close instance removal does

not compromise on the model performance, rather on the contrary, it improves the accuracy. This

simultaneous improvement of model fairness and accuracy which are in contrast of each other,

although seems to be astonishing, but we could provide rationale for this achievement using prior

work of Frénay et. al. [16]. This prior art explains the affect of such noisy instances on model

performance and claim that label noise hampers the performance of the classiﬁer which is also backed

by (Long 2008)[15].

To summarize, in this paper, we make the following contributions :

We propose a systematic way of identiﬁcation of bias inducing instances as per our deﬁnition

in the previous section, and their subsequent removal from training data.

We show that how the bias inducing instances removal ensures model fairness using the

standard fairness metric.

Further we show improvement in model accuracy trained on the bias eliminated data along

with justiﬁcation.

We offer control in terms of adjustable hyperparameters to adjust fairness and accuracy as

per the dataset and business requirements.

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：

SimultaneousImprovementofMLModelFairnessandPerformancebyIdentifyingBiasinDataAakashAgarwal,BhushanChaudhari,Dr.TanmoyBhowmikAIGarageMastercard(aakash.agarwal,bhushan.chaudhari,tanmoy.bhowmik)@mastercard.comAbstractMachinelearningmodelsbuiltondatasetscontainingdiscriminativeinstancesattributedtovario...

展开>> 收起<<

Simultaneous Improvement of ML Model Fairness and Performance by Identifying Bias in Data Aakash Agarwal Bhushan Chaudhari Dr. Tanmoy Bhowmik.pdf

共6页,预览2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！

Simultaneous Improvement of ML Model Fairness and Performance by Identifying Bias in Data Aakash Agarwal Bhushan Chaudhari Dr. Tanmoy Bhowmik

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: