A FEW-SHOT LEARNING APPROACH WITH DOMAIN ADAPTATION FOR PERSONALIZED REAL-LIFE STRESS DETECTION IN CLOSE RELATIONSHIPS Kexin Feng Jacqueline B. Duongy Kayla E. Cartay Sierra Waltersy

2025-04-30 0 0 335.83KB 5 页 10玖币

侵权投诉

A FEW-SHOT LEARNING APPROACH WITH DOMAIN ADAPTATION FOR

PERSONALIZED REAL-LIFE STRESS DETECTION IN CLOSE RELATIONSHIPS

Kexin Feng∗, Jacqueline B. Duong†, Kayla E. Carta†, Sierra Walters†,

Gayla Margolin§, Adela C. Timmons†, Theodora Chaspari∗

∗Texas A&M University, College Station, TX, USA

§University of Southern California, Los Angeles, California, USA

†University of Texas, Austin, TX, USA

∗{kexin, chaspari}@tamu.edu, §margolin@usc.edu

†{jduong, kcarta, snwalters}@utexas.edu, †adela.timmons@austin.utexas.edu

ABSTRACT

We design a metric learning approach that aims to address

computational challenges that yield from modeling human

outcomes from ambulatory real-life data. The proposed met-

ric learning is based on a Siamese neural network (SNN) that

learns the relative difference between pairs of samples from

a target user and non-target users, thus being able to address

the scarcity of labelled data from the target. The SNN fur-

ther minimizes the Wasserstein distance of the learned em-

beddings between target and non-target users, thus mitigat-

ing the distribution mismatch between the two. Finally, given

the fact that the base rate of focal behaviors is different per

user, the proposed method approximates the focal base rate

based on labelled samples that lay closest to the target, based

on which further minimizes the Wasserstein distance. Our

method is exempliﬁed for the purpose of hourly stress classi-

ﬁcation using real-life multimodal data from 72 dating cou-

ples. Results in few-shot and one-shot learning experiments

indicate that proposed formulation beneﬁts stress classiﬁca-

tion and can help mitigate the aforementioned challenges.

Index Terms—Few-shot learning, Siamese neural net-

work, Wasserstein distance, stress detection

1. INTRODUCTION

The increasing availability of Internet of Things (IoT) tech-

nologies and wearable devices has enabled the monitoring

of human states outside the lab resulting in the acquisition

of real-life, multimodal, temporal data, that can serve as the

foundation for automated algorithms for tracking an individ-

ual’s internal and contextual states. Detecting (or predict-

ing) one’s changing state can contribute to behavioral sup-

port delivery by providing the right type and amount of feed-

This research is based on work supported by NSF BCS-1627272 (Mar-

golin, PI), NIH NIMH R42MH123368 (Timmons, PI), and NSF IIS-2046118

(Chaspari, PI). A. C. Timmons owns intellectual property and stock in Col-

liga Apps Corp. and could beneﬁt ﬁnancially from commercialization of re-

lated research. The code of this work is available at: https://github.

com/HUBBS-Lab-TAMU/couple-stress-detection.

back at the right time [1]. Particularly, the prevalence of pro-

longed psychological stress globally [2] renders such appli-

cations more relevant than ever. Beyond the individual level,

ambulatory monitoring can be considered at an interpersonal

context for detecting stress between interacting partners in

close relationships (i.e., romantic couples, families) [3]. In

this context, ambulatory data can be used to monitor intercon-

nected stress patterns between the partners, thus contributing

to behavioral interventions that can demonstrate pathways for

handling stress at the interpersonal level [4, 5].

Designing machine learning (ML) models using real-

world human-centered data presents unique computational

challenges. First, labels obtained from the ambulatory

data may be costly and tend to depict low temporal reso-

lution [6]. In particular, labels in ambulatory monitoring are

commonly acquired via ecological momentary assessment

(EMA), which is usually not administered exactly at the time

of the stressor stimuli. While third-party annotation can po-

tentially ﬁx this issue, such type of annotation is difﬁcult

(or non-feasible) to obtain, since it requires human experts

reviewing large portions of longitudinal data. Second, due

to the inherent inter-individual variability, the distribution of

features that result from ambulatory data may be different

among individuals [7]. Due to this domain mismatch, the per-

formance of the models may vary across participants. Third,

the base rate of occurrence of the focal behaviors is usually

unknown and may vary among users [8].

To address the above challenges, we propose a distance

learning method implemented with a Siamese neural network

(SNN) that learns stress embeddings between labelled sam-

ples from non-target participants and a few labelled samples

from a target participant. We also integrate into the proposed

approach a distance-based criterion based on the Wasserstein

distance that aims to minimize the mismatch between the un-

labelled samples of the target participant and the rest of the

participants. Finally, to address the challenge of unknown

stress base rate, we compute the Wasserstein distance across

various hypothesized stress base rates for the target partic-

arXiv:2210.15247v1 [cs.LG] 27 Oct 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：

AFEW-SHOTLEARNINGAPPROACHWITHDOMAINADAPTATIONFORPERSONALIZEDREAL-LIFESTRESSDETECTIONINCLOSERELATIONSHIPSKexinFeng,JacquelineB.Duongy,KaylaE.Cartay,SierraWaltersy,GaylaMargolinx,AdelaC.Timmonsy,TheodoraChaspariTexasA&MUniversity,CollegeStation,TX,USAxUniversityofSouthernCalifornia,LosAngeles,Calif...

展开>> 收起<<

A FEW-SHOT LEARNING APPROACH WITH DOMAIN ADAPTATION FOR PERSONALIZED REAL-LIFE STRESS DETECTION IN CLOSE RELATIONSHIPS Kexin Feng Jacqueline B. Duongy Kayla E. Cartay Sierra Waltersy.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！

A FEW-SHOT LEARNING APPROACH WITH DOMAIN ADAPTATION FOR PERSONALIZED REAL-LIFE STRESS DETECTION IN CLOSE RELATIONSHIPS Kexin Feng Jacqueline B. Duongy Kayla E. Cartay Sierra Waltersy

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: