Outlier-Aware Training for Improving Group Accuracy Disparities Li-Kuang Chen1Canasai Kruengkrai2Junichi Yamagishi2 1National Tsing Hua University Taiwan

2025-04-29 0 0 343.57KB 7 页 10玖币

侵权投诉

Outlier-Aware Training for Improving Group Accuracy Disparities

Li-Kuang Chen1∗Canasai Kruengkrai2Junichi Yamagishi2

1National Tsing Hua University, Taiwan

lkchen@nlplab.cc

2National Institute of Informatics, Japan

{canasai,jyamagishi}@nii.ac.jp

Abstract

Methods addressing spurious correlations such

as Just Train Twice (JTT,Liu et al. 2021) in-

volve reweighting a subset of the training set to

maximize the worst-group accuracy. However,

the reweighted set of examples may potentially

contain unlearnable examples that hamper the

model’s learning. We propose mitigating this

by detecting outliers to the training set and re-

moving them before reweighting. Our experi-

ments show that our method achieves competi-

tive or better accuracy compared with JTT and

can detect and remove annotation errors in the

subset being reweighted in JTT.1

1 Introduction

Machine learning models trained with empirical

risk minimization (ERM, Vapnik 1992) can achieve

a high average accuracy by minimizing the over-

all loss during training. Despite this, ERM mod-

els are also known to perform poorly on certain

minority groups of examples. When speciﬁc at-

tributes in a dataset frequently co-occur with a

class label, ERM models often learn to correlate

the co-occurring attributes and the label, using the

attributes as “shortcuts” for classifying examples.

These “shortcuts” are also called spurious correla-

tions, because model performance can signiﬁcantly

decrease when the model encounters examples that

belong to a minority group where the correlations

between the attributes and class label do not hold.

More speciﬁcally, each class in a dataset can be

divided by whether their examples contain such

spurious attributes. Each set of examples with a

class-attribute combination is called a “group”. The

worst group is characterized by having the poor-

est ERM model performance among other groups.

As an example, Figure 1shows accuracy dispar-

ities among groups in the FEVER dataset. The

∗

This work was conducted during the author’s internship

under National Institute of Informatics, Japan.

Our code is available at

https://github.com/

nii-yamagishilab/jtt-m.

Figure 1: Results for the FEVER test set (Thorne et al.,

2018;Schuster et al.,2021). The data are divided into

six groups in accordance with class-attribute combi-

nations, where class = {REFUTES (REF), SUPPORTS

(SUP), NOT ENOUGH INFO (NEI)} and attribute = {no

neg, neg}, indicating the presence of a negation word in

the claim. Both methods perform well on groups with

strong spurious correlations (e.g., [REF, neg]). Our pro-

posed method (JTT-m) helps improve accuracies for

groups where such spurious correlations do not appear

(e.g., [SUP, neg] and [NEI, neg]).

ERM-trained model can achieve close to perfect

accuracy on the group with a spurious correlation

(the REFUTES class with negation), but only half

the accuracy on the worst group (the SUPPORTS

class with negation).

Improving the worst-group performance of ERM

models while maintaining the overall accuracy is

an active topic of research that has applications

in fair machine learning classiﬁers or robustness

against adversarial examples (Słowik and Bottou,

2022). Methods aiming to maximize worst-group

accuracy can be roughly categorized into two cat-

egories: those that utilize group information and

those that do not. Group Distributionally Robust

Optimization (Group DRO, Sagawa et al. 2020)

uses attribute (and thus group) information during

training to dynamically minimize the loss of each

group. While Group DRO achieves a high worst-

group and overall accuracy, it requires annotation

arXiv:2210.15183v1 [cs.CL] 27 Oct 2022

on group information during training, which can

be expensive to obtain and unavailable for less pop-

ular datasets. On the other hand, methods such

as DRO with Conditional Value-at-Risk (CVaR

DRO, Duchi et al. 2019;Levy et al. 2020), Learn-

ing from Failure (LfF, Nam et al. 2020), Predict

then Interpolate (PI, Bao et al. 2021), Spectral De-

coupling (SD, Pezeshki et al. 2021), Just Train

Twice (

JTT

,Liu et al. 2021), and RWY and SUBY

from (Idrissi et al.,2022) all aim to minimize worst-

group loss without group information.

CVaR DRO minimizes worst-case loss over all

subpopulations of a speciﬁc size and requires com-

puting the worst-case loss at each step. LfF trains

an intentionally biased model and upweights the

minority examples. PI interpolates distributions

of correct and incorrect predictions and can min-

imize worst-case loss over all interpolations. SD

replaces the L

weight decay in the cross entropy

loss function with logits. RWY reweights sam-

pling probabilities so that mini-batches are class-

balanced. SUBY subsamples large classes so that

every class is the same size as the smallest class.

JTT

simply obtains misclassiﬁed examples (the er-

ror set) from the training set once and upweights

the ﬁxed set of erroneous examples. We focus on

JTT

due to its simplicity and relative effectiveness

and because it does not require group information

for improving worst-group accuracy. While Idrissi

et al. (2022)’s SUBY and RWY also follow

JTT

improving worst-group accuracies, their methods

target only datasets with imbalanced classes, and

are not applicable to class-balanced datasets such

as MultiNLI (Williams et al.,2018).

We propose further enhancing

JTT

by removing

outliers from the error set before upweighting it.

The outliers might be examples that are difﬁcult

to learn, such as annotation errors. Keeping them

from being upweighted allows the model to train on

a cleaner error set and thus better show the intended

effect of the original

JTT

. We focus on worst-group

performance caused by the spurious correlations

of negation words and evaluate on datasets sus-

ceptible to spurious correlations of this type. Our

experiments on the FEVER and MultiNLI datasets

show that our method can outperform

JTT

in terms

of either the average or the worst-group accuracy

while maintaining the same level of performance

for the other groups.

Our contributions are as follows. We devise a

method for improving worst-group accuracy with-

out group information during training based on

JTT

(Section 3). We show that by removing out-

liers from the error set being upweighted, we can

achieve similar or better overall and worst-group

performance (Section 4.2). Our examination of the

outliers being removed also suggests that the im-

provement may come from removing annotation

errors in the upweighted error set (Section 4.3).

2 Background

Spurious correlations and minority groups

We investigate the spurious correlations occurring

in two natural-language datasets: FEVER (Thorne

et al.,2018) and MultiNLI (Williams et al.,2018).

The task for FEVER involves retrieving docu-

ments related to a given claim, ﬁnding sentences

to form evidence against the claim, and then clas-

sifying the claim on the basis of the evidence into

three classes: SUPPORTS (SUP), REFUTES (REF),

or NOT ENOUGH INFORMATION (NEI). We fo-

cus on improving the worst-group classiﬁcation

performance for the ﬁnal part of the task. The

task for MultiNLI is to classify whether the hy-

pothesis is entailed by, neutral with, or contra-

dicted by the premise. We use Schuster et al.

(2021)’s preprocessing of both datasets, contain-

ing 178,059/11,620/11,710 training/dev/test exam-

ples for FEVER and 392,702/9,832 training/test

examples for MultiNLI.

Attributes known to cause spurious correlations

for these datasets are negation words (Gururangan

et al.,2018) and verbs that suggest negating actions

(Schuster et al.,2019). We merge these two sources

of negation words into a single set: {no, never,

nothing, nobody, not, yet, refuse, refuses, refused,

fail, fails, failed, only, incapable, unable, neither,

none}. Each class can be split into two groups

based on whether each claim/hypothesis contains

a spurious attribute (i.e., the negation words listed

above). Models tend to perform well on groups

where the attributes are highly correlated with the

label. Groups where the correlation between the

label and the attribute does not hold are called mi-

nority groups or worst groups, since models often

fail to classify their examples correctly. For exam-

ple, the claim “Luis Fonsi does

not

go by his given

name on stage.”, labeled SUPPORTS, belongs to the

worst group [SUP, neg].

Table 1(a) shows that most claims containing

negation are from the class REFUTES. The rela-

tively small amount of examples from the groups

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：

Outlier-AwareTrainingforImprovingGroupAccuracyDisparitiesLi-KuangChen1CanasaiKruengkrai2JunichiYamagishi21NationalTsingHuaUniversity,Taiwanlkchen@nlplab.cc2NationalInstituteofInformatics,Japan{canasai,jyamagishi}@nii.ac.jpAbstractMethodsaddressingspuriouscorrelationssuchasJustTrainTwice(JTT,Liuetal...

展开>> 收起<<

Outlier-Aware Training for Improving Group Accuracy Disparities Li-Kuang Chen1Canasai Kruengkrai2Junichi Yamagishi2 1National Tsing Hua University Taiwan.pdf

共7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！

Outlier-Aware Training for Improving Group Accuracy Disparities Li-Kuang Chen1Canasai Kruengkrai2Junichi Yamagishi2 1National Tsing Hua University Taiwan

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: