A Practical Divide-and-Conquer Approach for Preference-Based Learning to Rank
Date Issued
2014
Date
2014
Author(s)
Jay, Yang-Han
Abstract
In preference-based learning to rank (LTR), rather than training a score- based prediction model, a binary prediction model (with probabilistic output) is trained over pairs of instances as a preference function. The ranking is then produced using the pairwise preference outputs in the prediction stage. In this paper we study the preference-based LTR problem and presents a practical approach we called the “Fuzzy Sort” which runs in O(W·N lg N), where W is typically no larger than 50 in practice. The algorithm shows promising results compared with other conventional ranking methods, and is query-efficient when competing against other preference-based LTR approaches.
Subjects
機器學習
排名學習
偏好
分治
Type
thesis
File(s)![Thumbnail Image]()
Loading...
Name
ntu-103-R01922064-1.pdf
Size
23.32 KB
Format
Adobe PDF
Checksum
(MD5):378e9de2dbbda1e555bc29e6a5fbeddd