Information Technology Reference
In-Depth Information
2.5.1 Relationship with Relevance Feedback
The pointwise approach to learning to rank, especially the classification-based algo-
rithms, has strong correlation with the relevance feedback algorithms [
7
,
19
]. The
relevance feedback algorithms, which have played an important role in the literature
of information retrieval, also leverage supervised learning technologies to improve
the retrieval accuracy. The basic idea is to learn from explicit, implicit, or blind feed-
back to update the original query. Then the new query is used to retrieve a new set
of documents. By doing this in an iterative manner, we can bring the original query
closer to the optimal query so as to improve the retrieval performance.
Here we take the famous Rocchio algorithm [
19
] as an example to make dis-
cussions on the relationship between relevance feedback and learning to rank. The
specific way that the Rocchio algorithm updates the query is as follows. First, both
the query
q
and its associated documents are represented in a vector space. Sec-
ond, through relevance feedback,
m
+
j
{
x
j
}
are identified as relevant documents (i.e.,
=
1
m
+
+
m
−
y
j
=
1), and
{
x
j
}
1
are identified as irrelevant documents (i.e.,
y
j
=
0). Third,
m
+
+
j
=
the query vector is updated according to the following heuristic:
m
+
m
+
+
m
−
1
m
+
1
m
−
q
˜
=
αq
+
β
x
j
−
γ
x
j
.
(2.17)
j
=
1
m
+
+
j
=
1
If we use the VSM model for retrieval, the documents will then be ranked ac-
cording to their inner product with the new query vector q. Mathematically, we can
define the corresponding scoring function as
q
T
x
j
.
f(x
j
)
=˜
(2.18)
In this sense, we can regard the new query vector as the model parameter. For
ease of discussion, we use
w
to represent this vector, i.e.,
w
q
. Then, as shown
in [
14
], there is actually a hidden loss function behind the above query update pro-
cess, which is a function of
w
and
x
. That is,
=˜
m
+
(
1
−
α
1
2
βw
T
x
j
),
w
−
y
j
=
1
,
4
L(f
;
x
j
,y
j
)
=
(2.19)
m
−
(
1
−
α
1
2
+
γw
T
x
j
),
w
y
j
=
0
.
4
In other words, the Rocchio algorithm also minimizes a certain pointwise loss
function. In this sense, it looks quite similar to the pointwise approach to learning
to rank. However, we would like to point out its significant differences from what
we call learning to rank, as shown below.
•
The feature space in the Rocchio algorithm is the standard vector space as used
in VSM. In this space, both query and documents are represented as vectors,
and their inner product defines the relevance. In contrast, in learning to rank, the
feature space contains features extracted from each query-document pair. Only
documents have feature representations, and the query is not a vector in the same
feature space.
