Information Technology Reference
In-Depth Information
Chapter 7
Query-Dependent Ranking
Abstract
In this chapter, we introduce query-dependent ranking. Considering the
large differences between different queries, it might not be the best choice to use a
single ranking function to deal with all kinds of queries. Instead, one may achieve
performance gain by leveraging the query differences. To consider the query differ-
ence in training, one can use a query-dependent loss function. To further consider the
query difference in the test process, a query-dependent ranking function is needed.
Several ways of learning a query-dependent ranking function are reviewed in this
chapter, including query classification-based approach, query clustering-based ap-
proach, nearest neighbor-based approach, and two-layer learning-based approach.
Discussions are also made on the future research directions along this line.
Queries in web search may vary largely in semantics and the users' intentions they
represent, in forms they appear, and in number of relevant documents they have in
the document repository. For example, queries can be navigational, informational,
or transactional. Queries can be personal names, product names, or terminology.
Queries can be phrases, combinations of phrases, or natural language sentences.
Queries can be short or long. Queries can be popular (which have many relevant
documents) or rare (which only have a few relevant documents). If one can success-
fully leverage query differences in the process of learning to rank, one may have
the opportunities to achieve better ranking performances in both training and test
processes.
7.1 Query-Dependent Loss Function
In [
12
], the query differences are considered in the training process. This is achieved
by using a query-dependent loss function. For ease of illustration, we assume that a
regression-based algorithm is used for learning to rank, whose original loss function
is defined as below.
m
(i)
n
y
(i)
j
f
x
(i)
j
2
2
.
L(f )
=
−
+
λ
f
f
(7.1)
i
=
j
=
1
1
