서브메뉴
검색
Towards Scalability and Robustness for Ranking, Clustering, and Multi-Armed Bandits
Towards Scalability and Robustness for Ranking, Clustering, and Multi-Armed Bandits
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20250211151007
- ISBN
- 9798382830643
- DDC
- 004
- 서명/저자
- Towards Scalability and Robustness for Ranking, Clustering, and Multi-Armed Bandits
- 발행사항
- [Sl] : University of Pennsylvania, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 237 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 85-12, Section: A.
- 주기사항
- Advisor: Khanna, Sanjeev;Agarwal, Shivani.
- 학위논문주기
- Thesis (Ph.D.)--University of Pennsylvania, 2024.
- 초록/해제
- 요약In recent years, machine learning has become an indispensable tool across various industry domains, revolutionizing the way businesses leverage data to make decisions. One of the key factors driving the success of machine learning is the explosion of big data which has enabled the development of increasingly sophisticated models. However, this era of big data has brought with it a unique set of challenges: prominently, (a) issues regarding data quality, and (b) scalability of learning algorithms. Addressing these challenges is essential in order to more effectively harness the synergy between big data and machine learning.This dissertation systematically addresses these challenges for certain foundational learning tasks: ranking from pairwise comparisons, hierarchical clustering, and online learning with bandit feedback. For each of these problems, we present an essentially-complete picture through algorithmic upper-bounds, and complementary lower-bounds for achieving adversarial robustness, high parallelizability and memory efficiency. More specifically,We study the problem of robust estimation for the popular BTL model for ranking from offline data in a very general adversarial contamination setting. In this setting, we establish exact necessary and sufficient conditions for identifiability, showing that robustness is a structural property of the underlying topology itself. We further show that a popular class of comparison graphs - Erdos-Renyi graphs - is highly robust to contamination. For these graphs, we also design an estimation algorithm that can provably recover from a non-trivial corruption rate.We then consider robust ranking in the online setting as a robust regret minimization problem for dueling bandits in a very general adversarial contamination setting. We design an algorithm that, assuming only the existence of a Condorcet winner, achieves low regret without any foreknowledge of the extent of corruption in the feedback. Moreover, we show that this regret is asymptotically optimal via a lower bound.On the scalability front, we consider the top-k identification problem for ranking in a limited-adaptivity setting. When the preferences satisfy strong stochastic transitivity, we establish a lower bound on the sample-complexity with unrestricted adaptivity, and show that just 3 adaptive rounds are sufficient to achieve it (up to log factors). We also design more general multi-round algorithms with even lower sample complexities, establishing a non-trivial upper-bound on the round vs sample complexity tradeoff.We then consider the problem of scalable minimum cost hierarchical clustering from paired similarity data, where we design highly parallel and memory-efficient algorithms in the massively parallel computing, and streaming model, respectively. Both algorithms are provably optimal, and follow as a consequence of novel structural properties that we prove about the clustering cost function itself. These properties and consequently, our algorithmic results, also extend to other maximization based objectives for this problem.Lastly, we consider scalability for classical stochastic bandits in a space-bounded multi-pass streaming setting where we uncover a surprising phenomenon: increasing memory beyond a constant to any quantity that is sublinear in the number of arms has almost no effect on reducing regret in the worst case. We design a constant-memory algorithm that achieves low regret - both worst-case and instance-dependent - in a given fixed number of passes, and prove a highly technical lower bound showing it is not possible to do much better in those many passes, even when allowed additional superconstant memory.
- 일반주제명
- Computer science
- 일반주제명
- Statistics
- 일반주제명
- Information science
- 키워드
- Machine learning
- 키워드
- Scalability
- 키워드
- Robustness
- 키워드
- Clustering
- 기타저자
- University of Pennsylvania Computer and Information Science
- 기본자료저록
- Dissertations Abstracts International. 85-12A.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008250123s2024 us c eng d■001000017160376
■00520250211151007
■006m o d
■007cr#unu||||||||
■020 ▼a9798382830643
■035 ▼a(MiAaPQ)AAI30994948
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a004
■1001 ▼aPatil, Prathamesh.
■24510▼aTowards Scalability and Robustness for Ranking, Clustering, and Multi-Armed Bandits
■260 ▼a[Sl]▼bUniversity of Pennsylvania▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a237 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 85-12, Section: A.
■500 ▼aAdvisor: Khanna, Sanjeev;Agarwal, Shivani.
■5021 ▼aThesis (Ph.D.)--University of Pennsylvania, 2024.
■520 ▼aIn recent years, machine learning has become an indispensable tool across various industry domains, revolutionizing the way businesses leverage data to make decisions. One of the key factors driving the success of machine learning is the explosion of big data which has enabled the development of increasingly sophisticated models. However, this era of big data has brought with it a unique set of challenges: prominently, (a) issues regarding data quality, and (b) scalability of learning algorithms. Addressing these challenges is essential in order to more effectively harness the synergy between big data and machine learning.This dissertation systematically addresses these challenges for certain foundational learning tasks: ranking from pairwise comparisons, hierarchical clustering, and online learning with bandit feedback. For each of these problems, we present an essentially-complete picture through algorithmic upper-bounds, and complementary lower-bounds for achieving adversarial robustness, high parallelizability and memory efficiency. More specifically,We study the problem of robust estimation for the popular BTL model for ranking from offline data in a very general adversarial contamination setting. In this setting, we establish exact necessary and sufficient conditions for identifiability, showing that robustness is a structural property of the underlying topology itself. We further show that a popular class of comparison graphs - Erdos-Renyi graphs - is highly robust to contamination. For these graphs, we also design an estimation algorithm that can provably recover from a non-trivial corruption rate.We then consider robust ranking in the online setting as a robust regret minimization problem for dueling bandits in a very general adversarial contamination setting. We design an algorithm that, assuming only the existence of a Condorcet winner, achieves low regret without any foreknowledge of the extent of corruption in the feedback. Moreover, we show that this regret is asymptotically optimal via a lower bound.On the scalability front, we consider the top-k identification problem for ranking in a limited-adaptivity setting. When the preferences satisfy strong stochastic transitivity, we establish a lower bound on the sample-complexity with unrestricted adaptivity, and show that just 3 adaptive rounds are sufficient to achieve it (up to log factors). We also design more general multi-round algorithms with even lower sample complexities, establishing a non-trivial upper-bound on the round vs sample complexity tradeoff.We then consider the problem of scalable minimum cost hierarchical clustering from paired similarity data, where we design highly parallel and memory-efficient algorithms in the massively parallel computing, and streaming model, respectively. Both algorithms are provably optimal, and follow as a consequence of novel structural properties that we prove about the clustering cost function itself. These properties and consequently, our algorithmic results, also extend to other maximization based objectives for this problem.Lastly, we consider scalability for classical stochastic bandits in a space-bounded multi-pass streaming setting where we uncover a surprising phenomenon: increasing memory beyond a constant to any quantity that is sublinear in the number of arms has almost no effect on reducing regret in the worst case. We design a constant-memory algorithm that achieves low regret - both worst-case and instance-dependent - in a given fixed number of passes, and prove a highly technical lower bound showing it is not possible to do much better in those many passes, even when allowed additional superconstant memory.
■590 ▼aSchool code: 0175.
■650 4▼aComputer science
■650 4▼aStatistics
■650 4▼aInformation science
■653 ▼aMachine learning
■653 ▼aScalability
■653 ▼aRobustness
■653 ▼aClustering
■653 ▼aMulti-armed bandits
■690 ▼a0984
■690 ▼a0723
■690 ▼a0800
■690 ▼a0463
■71020▼aUniversity of Pennsylvania▼bComputer and Information Science.
■7730 ▼tDissertations Abstracts International▼g85-12A.
■790 ▼a0175
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17160376▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


