본문

서브메뉴

Fast and Scalable Solvers for Penalized Regression with Sparsity
Fast and Scalable Solvers for Penalized Regression with Sparsity
Fast and Scalable Solvers for Penalized Regression with Sparsity

Detailed Information

자료유형  
 학위논문 서양
최종처리일시  
20260202103136
ISBN  
9798311957526
DDC  
621.31
저자명  
Yang, Jae Hyun.
서명/저자  
Fast and Scalable Solvers for Penalized Regression with Sparsity
발행사항  
[Sl] : Stanford University, 2025
발행사항  
Ann Arbor : ProQuest Dissertations & Theses, 2025
형태사항  
129 p
주기사항  
Source: Dissertations Abstracts International, Volume: 86-12, Section: B.
주기사항  
Advisor: Hastie, Trevor.
학위논문주기  
Thesis (Ph.D.)--Stanford University, 2025.
초록/해제  
요약In the last few decades, the landscape of machine learning changed drastically with the exponential growth of data in both the number of samples n and features p. Most regression models - the prototypical one being ordinary least squares (OLS) - favor the increase in n as they tend to generalize better. Moreover, the computational cost of large n can be easily ameliorated via distributed computing. However, as p grows, they typically suffer from overfitting, large memory consumptions, and increased computational complexity. The standard approach to tackle the overfitting issue is regularization. Many different regularizations exist today and some notable ones include the ridge, lasso [Friedman et al., 2010], elastic net [Zou and Hastie, 2005], and group lasso [Yuan and Lin, 2006]. The ridge adds the ℓ2penalty ∥β22 to the regression objective while the lasso adds the ℓ1penalty ∥β∥1. The ridge solution typically remains dense, however, the lasso often enjoys sparsity in the solution due to the soft-thresholding property. For this reason, the lasso has gained wide popularity in high-dimensional regression settings where p can be extremely large (sometimes on the order of millions) as the state-of-the-art solvers bypass the aforementioned computational challenges by leveraging this sparsity property. The elastic net is a convex combination of the ridge and the lasso, allowing for a more flexible model. The group lasso adds the (sum of) ℓ2-norm penalties PG g=1 ∥βg∥2 where groups of coefficients βg are separately penalized. This model tends to have group-wise sparse solutions (but typically dense within each non-zero group of coefficients).While the ridge, lasso, and elastic net have been studied carefully both theoretically and empirically, the same cannot be said about the group lasso. We believe a large reason for this phenomenon is the inexistence of a gold-standard package like glmnet [Friedman et al., 2010] for solving the group lasso. In Chapter 2, we propose a novel algorithm to solve the group lasso extremely efficiently and provide an industry-grade package, adelie, available in both Python and R. Our main contribution lies in solving the block update in the block-coordinate descent algorithm, which achieves a quadratic rate of convergence. Moreover, we solve the group lasso problem in full generality, making no assumptions about the structure of the feature matrix Xt hat other solvers impose. Many of the tricks employed in glmnet are naturally carried over to our solver such as warm-starts, active set strategy, and extensions to general convex losses via the proximal quasi-Newton method. In Chapter 3, we propose a heuristic called the pivot rule for discarding features in our algorithm. We often observe extremely fast convergence with the pivot rule especially when p is large, comparable to the state-of-the-art strong rule and significantly faster when there exists high pairwise correlation in the features. The pivot rule is largely motivated by the strong rule and only uses the KKT conditions to derive a robust ranking system of the features to guess the next active features. We use a simple change-point detection algorithm via OLS to automatically make our selection.
일반주제명  
Electricity
일반주제명  
Visualization
일반주제명  
Computer science
기타저자  
Stanford University.
기본자료저록  
Dissertations Abstracts International. 86-12B.
전자적 위치 및 접속  
로그인 후 원문을 볼 수 있습니다.

MARC

 008260126s2025        us                              c    eng  d
■001000017357130
■00520260202103136
■006m          o    d                
■007cr#unu||||||||
■020    ▼a9798311957526
■035    ▼a(MiAaPQ)AAI31974610
■035    ▼a(MiAaPQ)Stanfordhh827kx2104
■040    ▼aMiAaPQ▼cMiAaPQ
■0820  ▼a621.31
■1001  ▼aYang,  Jae  Hyun.
■24510▼aFast  and  Scalable  Solvers  for  Penalized  Regression  with  Sparsity
■260    ▼a[Sl]▼bStanford  University▼c2025
■260  1▼aAnn  Arbor▼bProQuest  Dissertations  &  Theses▼c2025
■300    ▼a129  p
■500    ▼aSource:  Dissertations  Abstracts  International,  Volume:  86-12,  Section:  B.
■500    ▼aAdvisor:  Hastie,  Trevor.
■5021  ▼aThesis  (Ph.D.)--Stanford  University,  2025.
■520    ▼aIn  the  last  few  decades,  the  landscape  of  machine  learning  changed  drastically  with  the  exponential  growth  of  data  in  both  the  number  of  samples  n  and  features  p.  Most  regression  models  -  the  prototypical  one  being  ordinary  least  squares  (OLS)  -  favor  the  increase  in  n  as  they  tend  to  generalize  better.  Moreover,  the  computational  cost  of  large  n  can  be  easily  ameliorated  via  distributed  computing.  However,  as  p  grows,  they  typically  suffer  from  overfitting,  large  memory  consumptions,  and  increased  computational  complexity.  The  standard  approach  to  tackle  the  overfitting  issue  is  regularization.  Many  different  regularizations  exist  today  and  some  notable  ones  include  the  ridge,  lasso  [Friedman  et  al.,  2010],  elastic  net  [Zou  and  Hastie,  2005],  and  group  lasso  [Yuan  and  Lin,  2006].  The  ridge  adds  the  ℓ2penalty  ∥β22  to  the  regression  objective  while  the  lasso  adds  the  ℓ1penalty  ∥β∥1.  The  ridge  solution  typically  remains  dense,  however,  the  lasso  often  enjoys  sparsity  in  the  solution  due  to  the  soft-thresholding  property.  For  this  reason,  the  lasso  has  gained  wide  popularity  in  high-dimensional  regression  settings  where  p  can  be  extremely  large  (sometimes  on  the  order  of  millions)  as  the  state-of-the-art  solvers  bypass  the  aforementioned  computational  challenges  by  leveraging  this  sparsity  property.  The  elastic  net  is  a  convex  combination  of  the  ridge  and  the  lasso,  allowing  for  a  more  flexible  model.  The  group  lasso  adds  the  (sum  of)  ℓ2-norm  penalties  PG  g=1  ∥βg∥2  where  groups  of  coefficients  βg  are  separately  penalized.  This  model  tends  to  have  group-wise  sparse  solutions  (but  typically  dense  within  each  non-zero  group  of  coefficients).While  the  ridge,  lasso,  and  elastic  net  have  been  studied  carefully  both  theoretically  and  empirically,  the  same  cannot  be  said  about  the  group  lasso.  We  believe  a  large  reason  for  this  phenomenon  is  the  inexistence  of  a  gold-standard  package  like  glmnet  [Friedman  et  al.,  2010]  for  solving  the  group  lasso.  In  Chapter  2,  we  propose  a  novel  algorithm  to  solve  the  group  lasso  extremely  efficiently  and  provide  an  industry-grade  package,  adelie,  available  in  both  Python  and  R.  Our  main  contribution  lies  in  solving  the  block  update  in  the  block-coordinate  descent  algorithm,  which  achieves  a  quadratic  rate  of  convergence.  Moreover,  we  solve  the  group  lasso  problem  in  full  generality,  making  no  assumptions  about  the  structure  of  the  feature  matrix  Xt  hat  other  solvers  impose.  Many  of  the  tricks  employed  in  glmnet  are  naturally  carried  over  to  our  solver  such  as  warm-starts,  active  set  strategy,  and  extensions  to  general  convex  losses  via  the  proximal  quasi-Newton  method.  In  Chapter  3,  we  propose  a  heuristic  called  the  pivot  rule  for  discarding  features  in  our  algorithm.  We  often  observe  extremely  fast  convergence  with  the  pivot  rule  especially  when  p  is  large,  comparable  to  the  state-of-the-art  strong  rule  and  significantly  faster  when  there  exists  high  pairwise  correlation  in  the  features.  The  pivot  rule  is  largely  motivated  by  the  strong  rule  and  only  uses  the  KKT  conditions  to  derive  a  robust  ranking  system  of  the  features  to  guess  the  next  active  features.  We  use  a  simple  change-point  detection  algorithm  via  OLS  to  automatically  make  our  selection.
■590    ▼aSchool  code:  0212.
■650  4▼aElectricity
■650  4▼aVisualization
■650  4▼aComputer  science
■690    ▼a0984
■71020▼aStanford  University.
■7730  ▼tDissertations  Abstracts  International▼g86-12B.
■790    ▼a0212
■791    ▼aPh.D.
■792    ▼a2025
■793    ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17357130▼nKERIS▼z이  자료의  원문은  한국교육학술정보원에서  제공합니다.

Preview

Export

ChatGPT Discussion

AI Recommended Related Books


    New Books MORE
    Statistics for the past 3 years. Go to brief

    Подробнее информация.

    • Бронирование
    • не существует
    • моя папка
    • Первый запрос зрения
    • Non-Book Loan Application
    • Nighttime Book Loan Application
    материал
    Reg No. Количество платежных Местоположение статус Ленд информации
    TF16807 전자도서 대출가능 My Folder 부재도서신고 비도서대출신청 야간 도서대출신청

    * Бронирование доступны в заимствований книги. Чтобы сделать предварительный заказ, пожалуйста, нажмите кнопку бронирование

    Books borrowed together with this book

    Related Popular Books

    Available after logging in.