서브메뉴
검색
Exploring Interpretable Latent Structure in Modern Data by Bayesian Modeling: Theory and Applications
Exploring Interpretable Latent Structure in Modern Data by Bayesian Modeling: Theory and Applications
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202105234
- ISBN
- 9798291567562
- DDC
- 310
- 서명/저자
- Exploring Interpretable Latent Structure in Modern Data by Bayesian Modeling: Theory and Applications
- 발행사항
- [Sl] : University of Michigan, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 397 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-03, Section: B.
- 주기사항
- Advisor: Nguyen, XuanLong.
- 학위논문주기
- Thesis (Ph.D.)--University of Michigan, 2025.
- 초록/해제
- 요약The increasing complexity of data in the modern era has necessitated the need for flexible statistical approaches, which require uncovering various types of latent structures in the data, often intricately linked to the heterogeneity in the population. This dissertation explores probabilistic models, including latent variable models and hierarchical models under a Bayesian framework in various settings, reinforcing their usefulness in extracting meaningful patterns and representations from data and broadening our understanding of such models both from theoretical and computational perspectives.The first chapter deals with developing models and theoretical understanding for hierarchical topic models, characterized by a latent tree-structured hierarchy among the topics and leading to an insightful structure formed by multiple topic polytopes sharing faces. Using such insight, this chapter explores the identifiability and contraction rates of the latent topics under the tree structure via geometric analysis under suitable asymptotic settings for group data.The second chapter builds on the geometric insights in the first chapter and extends to the case of continuous convolutional kernels. In particular, the results shed light on identifiability in general nonparametric mixtures of such convolutional distributions, where each component is supported nearly on a low-dimensional affine subspace. Novel inverse bound techniques allow characterizing posterior contraction rates in a generic parametrization of such models, under mild geometric assumptions, generalizing some classes of latent variable models considered in the existing literature.In the third chapter, we return to topic models, but in a different perspective. This chapter explores connections between the popular Latent Dirichlet Allocation and mixture of product multinomial models, using tensor decomposition of the Dirichlet distribution. Although this approach depends crucially on the admixing distribution, it enables a finer and more complete treatment of identifiability and posterior contraction rates than what is currently available in the existing literature. More generally, this provides a new avenue for studying hierarchical models by exploiting their correspondence with finite mixtures. The fourth chapter dives deeper into general hierarchical models under a grouped-data setting and extends the strong identifiability theory in mixture models to such models, by establishing appropriate inverse bounds for a range of asymptotic regimes. Instead of relying on the properties of the admixing distribution, in this chapter, we exploit the good properties associated with strongly identifiable probability kernels. The fifth chapter develops a non-parametric spatio-temporal model for dynamic velocity fields and focuses on scalable inference. Finally, the last chapter deals with contextual bandits under the Bayesian method, and by analyzing the posterior under such a complex environment, we provide regret guarantees for Thompson sampling under a sparse context setting.A recurring theme is the identifiability and posterior analysis of the model parameters. For latent variable models, often with complex dependence across the various parts of the model, studying parameter learning demands substantially more effort compared to density estimation properties; nevertheless, this understanding not only provides interpretability to such models but also aids in analyzing the performance of downstream tasks when using these models as part of the pipeline. Overall, this dissertation aims to deepen our understanding of some of the complex latent variable models commonly used in a variety of application domains and promotes the potential of such models for extracting interpretable structures from complex datasets.
- 일반주제명
- Statistics
- 일반주제명
- Computer science
- 기타저자
- University of Michigan Statistics
- 기본자료저록
- Dissertations Abstracts International. 87-03B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017359907
■00520260202105234
■006m o d
■007cr#unu||||||||
■020 ▼a9798291567562
■035 ▼a(MiAaPQ)AAI32271926
■035 ▼a(MiAaPQ)umichrackham006487
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a310
■1001 ▼aChakraborty, Sunrit.
■24510▼aExploring Interpretable Latent Structure in Modern Data by Bayesian Modeling: Theory and Applications
■260 ▼a[Sl]▼bUniversity of Michigan▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a397 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-03, Section: B.
■500 ▼aAdvisor: Nguyen, XuanLong.
■5021 ▼aThesis (Ph.D.)--University of Michigan, 2025.
■520 ▼aThe increasing complexity of data in the modern era has necessitated the need for flexible statistical approaches, which require uncovering various types of latent structures in the data, often intricately linked to the heterogeneity in the population. This dissertation explores probabilistic models, including latent variable models and hierarchical models under a Bayesian framework in various settings, reinforcing their usefulness in extracting meaningful patterns and representations from data and broadening our understanding of such models both from theoretical and computational perspectives.The first chapter deals with developing models and theoretical understanding for hierarchical topic models, characterized by a latent tree-structured hierarchy among the topics and leading to an insightful structure formed by multiple topic polytopes sharing faces. Using such insight, this chapter explores the identifiability and contraction rates of the latent topics under the tree structure via geometric analysis under suitable asymptotic settings for group data.The second chapter builds on the geometric insights in the first chapter and extends to the case of continuous convolutional kernels. In particular, the results shed light on identifiability in general nonparametric mixtures of such convolutional distributions, where each component is supported nearly on a low-dimensional affine subspace. Novel inverse bound techniques allow characterizing posterior contraction rates in a generic parametrization of such models, under mild geometric assumptions, generalizing some classes of latent variable models considered in the existing literature.In the third chapter, we return to topic models, but in a different perspective. This chapter explores connections between the popular Latent Dirichlet Allocation and mixture of product multinomial models, using tensor decomposition of the Dirichlet distribution. Although this approach depends crucially on the admixing distribution, it enables a finer and more complete treatment of identifiability and posterior contraction rates than what is currently available in the existing literature. More generally, this provides a new avenue for studying hierarchical models by exploiting their correspondence with finite mixtures. The fourth chapter dives deeper into general hierarchical models under a grouped-data setting and extends the strong identifiability theory in mixture models to such models, by establishing appropriate inverse bounds for a range of asymptotic regimes. Instead of relying on the properties of the admixing distribution, in this chapter, we exploit the good properties associated with strongly identifiable probability kernels. The fifth chapter develops a non-parametric spatio-temporal model for dynamic velocity fields and focuses on scalable inference. Finally, the last chapter deals with contextual bandits under the Bayesian method, and by analyzing the posterior under such a complex environment, we provide regret guarantees for Thompson sampling under a sparse context setting.A recurring theme is the identifiability and posterior analysis of the model parameters. For latent variable models, often with complex dependence across the various parts of the model, studying parameter learning demands substantially more effort compared to density estimation properties; nevertheless, this understanding not only provides interpretability to such models but also aids in analyzing the performance of downstream tasks when using these models as part of the pipeline. Overall, this dissertation aims to deepen our understanding of some of the complex latent variable models commonly used in a variety of application domains and promotes the potential of such models for extracting interpretable structures from complex datasets.
■590 ▼aSchool code: 0127.
■650 4▼aStatistics
■650 4▼aComputer science
■653 ▼aHierarchical topic model
■653 ▼aMixed membership mixture model
■653 ▼aPosterior contraction theory
■653 ▼aNonparametric mixtures
■653 ▼aLatent Dirichlet Allocation
■653 ▼aThompson sampling
■690 ▼a0463
■690 ▼a0796
■690 ▼a0984
■71020▼aUniversity of Michigan▼bStatistics.
■7730 ▼tDissertations Abstracts International▼g87-03B.
■790 ▼a0127
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17359907▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


