Studies on Ordinal Ranking with Regression
Date Issued
2011
Date
2011
Author(s)
Ruan, Yu-Xun
Abstract
Ranking is a popular research problem in recent years and has been used in wide range of applications including
web-search engines and recommendation systems.
In this thesis, we study on two ranking problems, the ordinal ranking problem and the top-rank problem.
We propose a novel ranking approach, cost-sensitive ordinal classification via regression (COCR), which respects the discrete nature of the ordinal ranks in real-world data sets. In particular, COCR applies a theoretically-sound reduction from ordinal to binary classification
and solves the binary classification sub-tasks with point-wise regression. Furthermore, COCR allows specifying mis-ranking costs to further boost the ranking performance.
We conduct experiments on two ranking problems respectively.
On the ordinal ranking problem, we compare different approaches based on decision trees.
The results show that the proposed COCR can perform better on many data sets when coupled with the appropriate cost.
Furthermore, on the top-rank problem, we derive the corresponding cost of a popular ranking criterion, Expected Reciprocal Rank (ERR), and plug the cost into the COCR approach. The resulting ERR-tuned COCR enjoys the benefits of both efficiency by using point-wise regression and top-rank prediction accuracy from the ERR criterion.
Evaluations on two large-scale data sets, including
``Yahoo! Learning to Rank Challenge'' and ``Microsoft Learning to Rank'', verify that some basic COCR settings outperform commonly-used regression approaches significantly. In addition, even better performance can often be achieved by the ERR-tuned COCR.
web-search engines and recommendation systems.
In this thesis, we study on two ranking problems, the ordinal ranking problem and the top-rank problem.
We propose a novel ranking approach, cost-sensitive ordinal classification via regression (COCR), which respects the discrete nature of the ordinal ranks in real-world data sets. In particular, COCR applies a theoretically-sound reduction from ordinal to binary classification
and solves the binary classification sub-tasks with point-wise regression. Furthermore, COCR allows specifying mis-ranking costs to further boost the ranking performance.
We conduct experiments on two ranking problems respectively.
On the ordinal ranking problem, we compare different approaches based on decision trees.
The results show that the proposed COCR can perform better on many data sets when coupled with the appropriate cost.
Furthermore, on the top-rank problem, we derive the corresponding cost of a popular ranking criterion, Expected Reciprocal Rank (ERR), and plug the cost into the COCR approach. The resulting ERR-tuned COCR enjoys the benefits of both efficiency by using point-wise regression and top-rank prediction accuracy from the ERR criterion.
Evaluations on two large-scale data sets, including
``Yahoo! Learning to Rank Challenge'' and ``Microsoft Learning to Rank'', verify that some basic COCR settings outperform commonly-used regression approaches significantly. In addition, even better performance can often be achieved by the ERR-tuned COCR.
Subjects
ordinal ranking
list-wise ranking
cost-sensitive
regression
Type
thesis
File(s)![Thumbnail Image]()
Loading...
Name
ntu-100-R98944042-1.pdf
Size
23.32 KB
Format
Adobe PDF
Checksum
(MD5):713218bcfed134f6221c07fa6620b72e