FairGen Fair Synthetic Data Generation Himanshu Chaudhary 1Bhushan Chaudhari 1Aakash Agarwal2Kamna Meena1Tanmoy Bhowmik3 Abstract

2025-05-06 0 0 215.1KB 5 页 10玖币

侵权投诉

FairGen: Fair Synthetic Data Generation

Himanshu Chaudhary * 1 Bhushan Chaudhari * 1 Aakash Agarwal 2Kamna Meena 1Tanmoy Bhowmik 3

Abstract

With the rising adoption of Machine Learning

across the domains like banking, pharmaceutical,

ed-tech, etc, it has become utmost important to

adopt responsible AI methods to ensure models

are not unfairly discriminating against any group.

Given the lack of clean training data, generative

adversarial techniques are preferred to generate

synthetic data with several state-of-the-art archi-

tectures readily available across various domains

from unstructured data such as text, images to

structured datasets modelling fraud detection and

many more. These techniques overcome several

challenges such as class imbalance, limited train-

ing data, restricted access to data due to privacy

issues. Existing work focusing on generating fair

data either works for a certain GAN architecture

or is very difﬁcult to tune across the GANs. In

this paper, we propose a pipeline to generate fairer

synthetic data independent of the GAN architec-

ture. The proposed paper utilizes a pre-processing

algorithm to identify and remove bias inducing

samples. In particular, we claim that while gen-

erating synthetic data most GANs amplify bias

present in the training data but by removing these

bias inducing samples, GANs essentially focuses

more on real informative samples. Our experi-

mental evaluation on two open-source datasets

demonstrates how the proposed pipeline is gener-

ating fair data along with improved performance

in some cases.

Equal contribution

Mastercard, India

Credgenics, India

GoTo Group, India. Correspondence to: Himanshu Chaud-

hary

himanshu.chaudhary@mastercard.com

, Bhushan

Chaudhari

bhushan.chaudhari@mastercard.com

, Aakash

Agarwal

aakash.agarwal1307@gmail.com

, Kamna Meena

mkamna14@gmail.com

, Tanmoy Bhowmik

tantan-

moy@gmail.com>.

Proceedings of the

39 th

International Conference on Machine

Learning, Baltimore, Maryland, USA, PMLR 162, 2022. Copy-

right 2022 by the author(s).

1. Introduction

Machine learning algorithms are being used ubiquitously in

the current world ranging from recommendation systems,

fraud detection, facial recognition to Autonomous driving.

These algorithms uses historical data to automate decision

making. However, these machine learning models have un-

intended bias towards protected group as demonstrated by

(Calders &

Zliobait

e,2013). Biasness comes in the model

due to the training data being biased itself. Training data

can have unfairness based on personal protected attributes

such as gender race, religion, etc. and this unfairness in

data doesn’t limit to only real world data but it also get

propagated to models which are built on such biased data.

Deep Learning models require huge training data to be ac-

curate. But due to limited training data, class imbalance

issues, Synthetic data have gained prominence as a replace-

ment to get additional high quality real world training data.

Generative Adversarial Networks (GANs) are one of the

most used popular way to generate synthetic data (Good-

fellow et al.,2014). GANs are generative models that has

two components: generator and discriminator. Model is

trained in such a way that generator fools discriminator and

generates synthetic data almost similar to real data. This

generated high quality synthetic data can be used for dif-

ferent predictive analysis in case where real world data is

limited (Choi et al.,2017). Health and ﬁnancial datasets are

particularly ripe for a synthetic approach. These ﬁelds are

highly restricted by privacy laws. Synthetic data can help

researchers get the data they need without violating people’s

privacy. For institutions in health, ﬁnancial sectors they can

not share their customers’ data directly. However if they

want to open source their data by adhering to all regulatory

rules for privacy of their customers, then they can create

synthetic data using real customer’s data. In doing so, the

real data will not get shared and the generated synthetic data

has the same distributional properties as of real data and this

can be used by researchers to create and do groundbreaking

research.

Existing GAN techniques generates biased data by amplify-

ing the bias present in the training data. (Gupta et al.,2021)

showed bias ampliﬁcation across various GAN architectures

including deferentially private generation schemes. This

paper motivated us to ﬁnd ways of reducing bias from the

training data itself.

arXiv:2210.13023v2 [cs.LG] 1 Dec 2022

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：

FairGen:FairSyntheticDataGenerationHimanshuChaudhary*1BhushanChaudhari*1AakashAgarwal2KamnaMeena1TanmoyBhowmik3AbstractWiththerisingadoptionofMachineLearningacrossthedomainslikebanking,pharmaceutical,ed-tech,etc,ithasbecomeutmostimportanttoadoptresponsibleAImethodstoensuremodelsarenotunfairlydiscrim...

展开>> 收起<<

FairGen Fair Synthetic Data Generation Himanshu Chaudhary 1Bhushan Chaudhari 1Aakash Agarwal2Kamna Meena1Tanmoy Bhowmik3 Abstract.pdf

共5页,预览1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！

FairGen Fair Synthetic Data Generation Himanshu Chaudhary 1Bhushan Chaudhari 1Aakash Agarwal2Kamna Meena1Tanmoy Bhowmik3 Abstract

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: