자료유형 | 학위논문 |
---|---|
서명/저자사항 | Essays in Applied Statistics and Machine Learning. |
개인저자 | Zhang, Ningshan. |
단체저자명 | New York University. Statistics. |
발행사항 | [S.l.]: New York University., 2019. |
발행사항 | Ann Arbor: ProQuest Dissertations & Theses, 2019. |
형태사항 | 184 p. |
기본자료 저록 | Dissertations Abstracts International 81-04B. Dissertation Abstract International |
ISBN | 9781687975393 |
학위논문주기 | Thesis (Ph.D.)--New York University, 2019. |
일반주기 |
Source: Dissertations Abstracts International, Volume: 81-04, Section: B.
Advisor: Simonoff, Jeffrey S. |
이용제한사항 | This item must not be sold to any third party vendors. |
요약 | In this dissertation, we look at three problems in applied statistics and machine learning. The first chapter considers the problem of fitting deeply nested hierarchical linear mixed models and its application to large scale recommender systems. We propose a recursive moment-based method for fitting hierarchical generalized linear mixed models of arbitrarily deep hierarchies. We show by simulations and a real world recommender system problem that, our proposed method is orders of magnitude faster than using off-the-shelf maximum likelihood procedures, while admitting comparable prediction performances.The second chapter examines the problem of joint modeling of longitudinal and time-to-event data via the latent class approach. Under the assumption that the longitudinal and time-to-event outcomes are independent conditioning on latent classes, we propose a nonparametric joint latent class modeling approach based on trees (JLCT). Simulation results as well as a real world example on the PAQUID dataset show that, the tree-based approach can be much faster than the most common parametric joint latent class modeling approach, the joint latent class model (JLCM). Furthermore, by using time-varying covariates in modeling survival risks and latent class memberships, JLCT can lead to a much more favorable prediction performance than JLCM, which is restricted to only using time-invariant covariates.In the last chapter, we study the multiple-source adaptation problem. We provide guarantees that there exists a distribution weighted combining rule that is robust with respect to any target mixture of the source distributions. These guarantees hold in the case where the conditional probabilities for the source domains are distinct, and where the loss function is the cross-entropy loss and the solution is normalized. Moreover, we give new algorithms for determining this robust combination solution for the cross-entropy loss and the squared loss. We report the results of a series of experiments with real-world datasets, where our algorithm outperforms competing approaches. |
일반주제명 | Statistics. |
언어 | 영어 |
바로가기 |
: 이 자료의 원문은 한국교육학술정보원에서 제공합니다. |