Information Technology Reference
In-Depth Information
7.2.4 Two-Layer Learning Approach
In [
6
], He and Liu argued that no matter whether query categorization, clustering,
or nearest neighbors is being used, one basically only pre-trains a finite number of
models and selects the most suitable one from them to rank the documents for a
test query. However, given the extremely large space of queries, this “finite-model”
solution might not work well for all the test queries. To solve the problem, they
propose a novel solution with an “infinite granularity”.
Specifically, the training process of query-dependent ranking is defined as the
following optimization problem:
n
L
ˆ
y
(i)
,
y
(i)
,
min
ν
i
=
1
f
w
(i)
,
x
(i)
,
y
(i)
s.t.
ˆ
=
sort
◦
g
ν,z
(i)
,
w
(i)
=
·
=
where
f(
)
is the scoring function whose parameter is
w
g(ν,z)
, in which
z
is the
query feature vector, and
ν
is the parameter of function
g
.
From the above formulation, it can be seen that there are two layers of learning
in the training process:
•
The first layer is named as the
document layer
. This layer corresponds to learn-
ing the parameter
w
in
f(w,
x
)
. In other words, one learns how to generate the
ranking scores for the documents by combining their document features. The
document-layer learning is the same as the learning process in conventional learn-
ing to rank.
•
The second layer is named as the
query layer
. This layer regards the parameter
w
as the target, and tries to generate
w
by combining query features, i.e.,
w
=
g(ν,z)
. In other words, one no longer regards
w
as a deterministic vector, but
a function of the query features. Then, for different queries, since their query
features are different, their corresponding ranking models will therefore become
different.
The test process is also different from conventional learning to rank. Given a test
query
q
, there is no pre-determined ranking model
w
that can be used to rank its
associated documents. Instead one needs to generate such a model on the fly. That
is, one first obtains the query features
z
, and then uses the learned parameter
ν
to
create a ranking model
w
=
g(ν,z)
. This model is then used to rank the documents
associated with
q
based on their document features.
As compared to other query-dependent ranking methods, this two-layer method
has the following properties.
•
In this approach, the way of using the query features are learned according to
the relevance label and loss functions for ranking. In contrast, the way of using
the query features in previous methods is either unsupervised or not according to
