Author's School

School of Engineering & Applied Science

Author's Department/Program

Computer Science and Engineering

Language

English (en)

Date of Award

January 2010

Degree Type

Thesis

Degree Name

Master of Arts (MA)

Chair and Committee

Kilian Weinberger

Abstract

Learning how to rank a set of objects relative to an user defined query has received much interest in the machine learning community during the past decade. In fact, there have been two recent competitions hosted by internationally prominent search companies to encourage research on ranking web site documents. Recent literature on learning to rank has focused on three approaches: point-wise, pair-wise, and list-wise. Many different kinds of classifiers, including boosted decision trees, neural networks, and SVMs have proven successful in the field. This thesis surveys traditional point-wise techniques that use regression trees for web-search ranking. The thesis contains empirical studies on Random Forests and Gradient Boosted Decision Trees, with novel augmentations to them on real world data sets. We also analyze how these point-wise techniques perform on new areas of research for web-search ranking: transfer learning and feature-cost aware models.

Comments

Permanent URL: http://dx.doi.org/10.7936/K7HH6H4B

Share

COinS