Information Technology Reference
In-Depth Information
Partly because of the central role of ranker in search engines, great attention
has been paid to the research and development of ranking technologies. Note that
ranking is also the central problem in many other information retrieval applications,
such as collaborative filtering [
30
], question answering [
3
,
71
,
79
,
83
], multimedia
retrieval [
86
,
87
], text summarization [
53
], and online advertising [
16
,
49
]. In this
topic, we will mainly take document retrieval in search as an example. Note that
even document retrieval is not a narrow task. There are many different ranking sce-
narios of interest for document retrieval. For example, sometimes we need to rank
documents purely according to their relevance with regards to the query. In some
other cases, we need to consider the relationships of similarity [
73
], website struc-
ture [
22
], and diversity [
91
] between documents in the ranking process. This is also
referred to as relational ranking [
61
].
To tackle the problem of document retrieval, many heuristic ranking models have
been proposed and used in the literature of information retrieval. Recently, given
the amount of potential training data available, it has become possible to leverage
machine learning technologies to build effective ranking models. Specifically, we
call those methods that learn how to combine predefined features for ranking by
means of discriminative learning “learning-to-rank” methods.
In recent years, learning to rank has become a very hot research direction in
information retrieval. First, a large number of learning-to-rank papers have been
published at top conferences on machine learning and information retrieval. Exam-
ples include [
4
,
7
-
10
,
15
,
18
,
20
,
21
,
25
,
26
,
31
,
37
,
44
,
47
,
55
,
58
,
62
,
68
,
72
,
76
,
82
,
88
,
89
]. There have been multiple sessions in recent SIGIR conferences ded-
icated for the topic on learning to rank. Second, benchmark datasets like LETOR
[
48
] have been released to facilitate the research on learning to rank.
5
Many re-
search papers on learning to rank have used these datasets for their experiments,
which make their results easy to compare. Third, several activities regarding learn-
ing to rank have been organized. For example, the workshop series on Learning to
Rank for Information Retrieval (2007-2009), the workshop on Beyond Binary Rel-
evance: Preferences, Diversity, and Set-Level Judgments (2008), and the workshop
on Redundancy, Diversity, and Interdependent Document Relevance (2009) have
been organized at SIGIR. A special issue on learning to rank has been organized at
Information Retrieval Journal (2009). In Table
1.1
, we have listed the major activ-
ities regarding learning to rank in recent years. Active participation of researchers
in these activities has demonstrated the continued interest from the research com-
munity on the topic of learning to rank. Fourth, learning to rank has also become
a key technology in the industry. Several major search engine companies are using
learning-to-rank technologies to train their ranking models.
6
When a research area comes to this stage, several questions as follows naturally
arise.
5
http://research.microsoft.com/~LETOR/
.
6
http://glinden.blogspot.com/2005/06/msn-search-and-learning-to-rank.html
,
http://www.ysearchblog.com/2008/07/09/boss-the-next-step-in-our-open-search-ecosystem/
.
