Exploring and Improving Crowdsourced Data Labeling for Mobile Activity Recognition

Mairittha,  Nattaya

WEKO3

lat lon distance

[[sub_check.contents]]

[[sub_radio.contents]]

Field does not validate

[[sub_attr.contents]]　

インデックスツリー

アイテム

行動認識機械学習データセット収集のためのクラウドソーシングの研究

https://doi.org/10.18997/00008417

名前 / ファイル	ライセンス	アクション
kou_k_526.pdf (15.2 MB)

アイテムタイプ

学位論文 = Thesis or Dissertation(1)

公開日

2021-07-30

資源タイプ

資源タイプ識別子

http://purl.org/coar/resource_type/c_db06

資源タイプ

doctoral thesis

タイトル

Exploring and Improving Crowdsourced Data Labeling for Mobile Activity Recognition

言語

タイトル

行動認識機械学習データセット収集のためのクラウドソーシングの研究

言語

eng

著者

Mairittha, Nattaya

抄録

内容記述タイプ

Abstract

内容記述

In this thesis, we propose novel methods to explore and improve crowdsourced data labeling for mobile activity recognition. This thesis concerns itself with the quality （i.e., the performance of a classiﬁcation model）, quantity （i.e., the number of data collected）, and motivation （i.e., the process that initiates and maintains goal-oriented behaviors） of participant contributions in mobile activity data collection studies. We focus on achieving high-quality and consistent ground-truth labeling and, particularly, on user feedback’s impact under diﬀerent conditions. Although prior works have used several techniques to improve activity recognition performance, diﬀerences to our approach exist in terms of the end goals, proposed method, and implementation. Many researchers commonly investigate post-data collection to increase activity recognition accuracy, such as implementing advanced machine learning algorithms to improve data quality or exploring several preprocessing ways to increase data quantity. However, utilizing post-data collection results is very diﬃcult and time-consuming due to dirty data challenges for most real-world situations. Unlike those commonly used in other literature, in this thesis, we aim to motivate and sustain user engagement during their on-going-self-labeling task to optimize activity recognition accuracy. The outline of the thesis is as follows: In chapter 1 and 2, we brieﬂy introduce the thesis work and literature review. In Chapter 3, we introduce novel gamiﬁed active learning and inaccuracy detection for crowdsourced data labeling for an activity recognition system （CrowdAct） using mobile sensing. We exploited active learning to address the lack of accurate information. We presented the integration of gamiﬁcation into active learning to overcome the lack of motivation and sustained engagement. We introduced an inaccuracy detection algorithm to minimize inaccurate data. In Chapter 4, we introduce a novel method to exploit on-device deep learning inference using a long short-term memory （LSTM）-based approach to alleviate the labeling eﬀort and ground truth data collection in activity recognition systems using smartphone sensors. The novel idea behind this is that estimated activities are used as feedback for motivating users to collect accurate activity labels. In Chapter 5, we introduce a novel on-device personalization for data labeling for an activity recognition system using mobile sensing. The key idea behind this system is that estimated activities personalized for a speciﬁc individual user can be used as feedback to motivate user contribution and improve data labeling quality. We exploited ﬁnetuning using a Deep Recurrent Neural Network （RNN） to address the lack of suﬃcient training data and minimize the need for training deep learning on mobile devices from scratch. We utilized a model pruning technique to reduce the computation cost of on-device personalization without aﬀecting the accuracy. Finally, we built a robust activity data labeling system by integrating the two techniques outlined above, allowing the mobile application to create a personalized experience for the user. To demonstrate the proposed methods’ capability and feasibility in realistic settings, we developed and deployed the systems to real-world settings such as crowdsourcing. For the process of data labeling, we challenged online and self-labeling scenarios using inertial smartphone sensors, such as accelerometers. We recruited diverse participants and con- ducted the experiments both in a laboratory setting and in a semi-natural setting. We also applied both manual labeling and the assistance of semi-automated labeling. Addition- ally, we gathered massive labeled training data in activity recognition using smartphone sensors and other information such as user demographics and engagement. Chapter 6 oﬀers a brief discussion of the thesis. In Chapter 7, we conclude the thesis with conclusion and some future work issues. We empirically evaluated these methods across various study goals such as machine learning and descriptive and inferential statistics. Our results indicated that this study enabled us to eﬀectively collect crowdsourced activity data. Our work revealed clear opportunities and challenges in combining human and mobile phone-based sensing techniques for researchers interested in studying human behavior in situ. Researchers and practitioners can apply our ﬁndings to improve recognition accuracy and reduce unreliable labels by human users, increase the total number of collected responses, as well as enhance participant motivation for activity data collection.

内容記述タイプ

TableOfContents

内容記述

1 Introduction||2 Related work||3 Achieving High-Quality Crowdsourced Datasets in Mobile Activity Recognition||4 On-Device Deep Learning Inference for Activity Data Collection||5 On-Device Deep Personalization for Activity Data Collection||6 Discussion||7 Conclusion

備考

内容記述タイプ

Other

内容記述

九州工業大学博士学位論文学位記番号：工博甲第526号学位授与年月日：令和3年6月28日

キーワード

主題Scheme

Other

主題

Activity Recognition

キーワード

主題Scheme

Other

主題

Machine Learning

キーワード

主題Scheme

Other

主題

Data Collection

キーワード

主題Scheme

Other

主題

Crowdsourcing

キーワード

主題Scheme

Other

主題

Mobile Computing

キーワード

主題Scheme

Other

主題

Active Learning

アドバイザー

井上, 創造

学位授与番号

甲第526号

学位名

博士（工学）

学位授与年月日

2021-06-28

学位授与機関

学位授与機関識別子Scheme

kakenhi

学位授与機関識別子

17104

学位授与機関名

九州工業大学

学位授与年度

内容記述タイプ

Other

内容記述

令和3年度

出版タイプ

VoR

出版タイプResource

http://purl.org/coar/version/c_970fb48d4fbd8a85

アクセス権

open access

アクセス権URI

http://purl.org/coar/access_right/c_abf2

ID登録

10.18997/00008417

ID登録タイプ

JaLC

戻る

views

See details

	Views

Versions

Ver.1

2023-05-15 12:35:24.050491

Show All versions

Cite as

Other

エクスポート

OAI-PMH

JPCOAR 2.0
JPCOAR 1.0
DublinCore
DDI

Other Formats

インデックスリンク

インデックスツリー

アイテム

行動認識機械学習データセット収集のためのクラウドソーシングの研究

× Mairittha, Nattaya

Versions

Share

Cite as

Other

エクスポート

コミュニティ

メニューを最小化

インデックスリンク

インデックスツリー

アイテム

行動認識機械学習データセット収集のためのクラウドソーシングの研究

× Mairittha, Nattaya

Versions

Share

Cite as

Other

エクスポート

コミュニティ