서브메뉴
검색
Theoretical Foundations and Applications of Integrated Learning Architectures for Graphs.
Theoretical Foundations and Applications of Integrated Learning Architectures for Graphs.
상세정보
- 자료유형
- 학위논문
- Control Number
- 0017162972
- International Standard Book Number
- 9798384342045
- Dewey Decimal Classification Number
- 658
- Main Entry-Personal Name
- Dax, Victoria Magdalena.
- Publication, Distribution, etc. (Imprint
- [S.l.] : Stanford University., 2024
- Publication, Distribution, etc. (Imprint
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- Physical Description
- 116 p.
- General Note
- Source: Dissertations Abstracts International, Volume: 86-03, Section: A.
- General Note
- Advisor: Kochenderfer, Mykel.
- Dissertation Note
- Thesis (Ph.D.)--Stanford University, 2024.
- Summary, Etc.
- 요약Graph Neural Networks (GNNs) have become important in the machine learning landscape because of their ability to model complex, structured data. This thesis presents approaches for blending GNNs with other deep learning methods, such as the decision-making capabilities of reinforcement learning (RL) or the generative abilities of variational auto-encoders (VAEs), to enhance the practical functionality of GNNs and to expand their applicability in various domains.The fundamental challenge addressed in this thesis is overcoming the inherent difficulties in integrating GNNs with other methodologies. GNNs excel in processing structured data but face issues like oversmoothing, particularly with increasing network depth. On the other hand, methods such as VAEs offer flexible generative abilities but have their own set of training and scalability challenges. And, while recurrent neural networks (RNNs) excel at processing temporal patterns, they introduce concerns of catastrophic forgetting and vanishing gradients. The solutions explored here involve novel combinations of these diverse techniques, aiming to leverage their strengths while mitigating their weaknesses.We start by exploring GNNs' theoretical properties, especially their generalization properties, before transitioning into practical applications. First, we demonstrate an enhancement of GNNs by integrating RNNs for advanced time-series predictions in interconnected systems. Then, we combine GNNs with variational auto-encoders (VAEs) to improve out-of-distribution generalizability and model interpretability through disentanglement of the embedding space in motion prediction tasks. We end by discussing using GNNs in deep RL techniques, specifically for combinatorial optimization tasks.
- Subject Added Entry-Topical Term
- Decision making.
- Subject Added Entry-Topical Term
- Neural networks.
- Subject Added Entry-Topical Term
- Web studies.
- Added Entry-Corporate Name
- Stanford University.
- Host Item Entry
- Dissertations Abstracts International. 86-03A.
- Electronic Location and Access
- 로그인을 한후 보실 수 있는 자료입니다.
- Control Number
- joongbu:657992
MARC
008250224s2024 us ||||||||||||||c||eng d■001000017162972
■00520250211152118
■006m o d
■007cr#unu||||||||
■020 ▼a9798384342045
■035 ▼a(MiAaPQ)AAI31460351
■035 ▼a(MiAaPQ)Stanfordsp191ym8296
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a658
■1001 ▼aDax, Victoria Magdalena.
■24510▼aTheoretical Foundations and Applications of Integrated Learning Architectures for Graphs.
■260 ▼a[S.l.]▼bStanford University. ▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a116 p.
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-03, Section: A.
■500 ▼aAdvisor: Kochenderfer, Mykel.
■5021 ▼aThesis (Ph.D.)--Stanford University, 2024.
■520 ▼aGraph Neural Networks (GNNs) have become important in the machine learning landscape because of their ability to model complex, structured data. This thesis presents approaches for blending GNNs with other deep learning methods, such as the decision-making capabilities of reinforcement learning (RL) or the generative abilities of variational auto-encoders (VAEs), to enhance the practical functionality of GNNs and to expand their applicability in various domains.The fundamental challenge addressed in this thesis is overcoming the inherent difficulties in integrating GNNs with other methodologies. GNNs excel in processing structured data but face issues like oversmoothing, particularly with increasing network depth. On the other hand, methods such as VAEs offer flexible generative abilities but have their own set of training and scalability challenges. And, while recurrent neural networks (RNNs) excel at processing temporal patterns, they introduce concerns of catastrophic forgetting and vanishing gradients. The solutions explored here involve novel combinations of these diverse techniques, aiming to leverage their strengths while mitigating their weaknesses.We start by exploring GNNs' theoretical properties, especially their generalization properties, before transitioning into practical applications. First, we demonstrate an enhancement of GNNs by integrating RNNs for advanced time-series predictions in interconnected systems. Then, we combine GNNs with variational auto-encoders (VAEs) to improve out-of-distribution generalizability and model interpretability through disentanglement of the embedding space in motion prediction tasks. We end by discussing using GNNs in deep RL techniques, specifically for combinatorial optimization tasks.
■590 ▼aSchool code: 0212.
■650 4▼aDecision making.
■650 4▼aNeural networks.
■650 4▼aWeb studies.
■690 ▼a0800
■690 ▼a0646
■71020▼aStanford University.
■7730 ▼tDissertations Abstracts International▼g86-03A.
■790 ▼a0212
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17162972▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
