서브메뉴
검색
Robust Inference and Interpretability in Complex Statistical Models: From Model Distillation to Targeted Learning
Robust Inference and Interpretability in Complex Statistical Models: From Model Distillation to Targeted Learning
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202103113
- ISBN
- 9798288861864
- DDC
- 574
- 저자명
- Zhou, Yunzhe.
- 서명/저자
- Robust Inference and Interpretability in Complex Statistical Models: From Model Distillation to Targeted Learning
- 발행사항
- [Sl] : University of California, Berkeley, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 171 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-01, Section: B.
- 주기사항
- Advisor: Hooker, Giles;van der Laan, Mark.
- 학위논문주기
- Thesis (Ph.D.)--University of California, Berkeley, 2025.
- 초록/해제
- 요약This dissertation introduces statistical methods aimed at ensuring robust inference and stable interpretability in complex models. A key challenge in modern statistical learning is the lack of transparency in predictive models, whose interpretations can be unreliable or inconsistent-particularly in the presence of finite-sample variability or model uncertainty. To address this, I develop a general framework for reproducible model distillation, which constructs interpretable student models from black-box predictors while providing statistical guarantees on the stability of the resulting explanations. In addition, the dissertation highlights the significant potential of applying targeted learning methods to non-causal estimands, with a focus on explainable AI and mathematical ecology. Targeted learning offers a robust framework for bias correction, uncertainty quantification, and model selection in these settings. Specifically, I investigate uncertainty quantification for variable importance using the targeted learning framework, improving inference in small-sample regimes. I also develop efficient estimators for long-term population dynamics by applying targeted maximum likelihood estimation (TMLE) to integral projection models in ecology. Furthermore, I extend the infinitesimal jackknife (IJ) to estimate covariances between model predictions, enabling principled comparisons across model classes and ensembles.
- 일반주제명
- Biostatistics
- 일반주제명
- Statistics
- 일반주제명
- Mathematics
- 키워드
- Robust inference
- 키워드
- Interpretability
- 기타저자
- University of California, Berkeley Biostatistics
- 기본자료저록
- Dissertations Abstracts International. 87-01B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017356994
■00520260202103113
■006m o d
■007cr#unu||||||||
■020 ▼a9798288861864
■035 ▼a(MiAaPQ)AAI31936565
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a574
■1001 ▼aZhou, Yunzhe.
■24510▼aRobust Inference and Interpretability in Complex Statistical Models: From Model Distillation to Targeted Learning
■260 ▼a[Sl]▼bUniversity of California, Berkeley▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a171 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-01, Section: B.
■500 ▼aAdvisor: Hooker, Giles;van der Laan, Mark.
■5021 ▼aThesis (Ph.D.)--University of California, Berkeley, 2025.
■520 ▼aThis dissertation introduces statistical methods aimed at ensuring robust inference and stable interpretability in complex models. A key challenge in modern statistical learning is the lack of transparency in predictive models, whose interpretations can be unreliable or inconsistent-particularly in the presence of finite-sample variability or model uncertainty. To address this, I develop a general framework for reproducible model distillation, which constructs interpretable student models from black-box predictors while providing statistical guarantees on the stability of the resulting explanations. In addition, the dissertation highlights the significant potential of applying targeted learning methods to non-causal estimands, with a focus on explainable AI and mathematical ecology. Targeted learning offers a robust framework for bias correction, uncertainty quantification, and model selection in these settings. Specifically, I investigate uncertainty quantification for variable importance using the targeted learning framework, improving inference in small-sample regimes. I also develop efficient estimators for long-term population dynamics by applying targeted maximum likelihood estimation (TMLE) to integral projection models in ecology. Furthermore, I extend the infinitesimal jackknife (IJ) to estimate covariances between model predictions, enabling principled comparisons across model classes and ensembles.
■590 ▼aSchool code: 0028.
■650 4▼aBiostatistics
■650 4▼aStatistics
■650 4▼aMathematics
■653 ▼aRobust inference
■653 ▼aInterpretability
■653 ▼aTargeted learning
■653 ▼aInfinitesimal jackknife
■653 ▼aTargeted maximum likelihood estimation
■690 ▼a0308
■690 ▼a0800
■690 ▼a0405
■690 ▼a0463
■71020▼aUniversity of California, Berkeley▼bBiostatistics.
■7730 ▼tDissertations Abstracts International▼g87-01B.
■790 ▼a0028
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17356994▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
