Information Technology Reference
In-Depth Information
4.4.2 Query and Image Language Models
In personalized search, the query model should consider both the query and user
information. Following the topic-sensitive influence modeling discussed in previous
sections, we define query model
u
ʸ
Q
as the distribution over the learned topics given
query
q
and user
u
:
p
(
z
|
u
)
p
(
q
|
z
,
u
)
u
ʸ
Q
p
(
z
|
q
,
u
)
=
(4.12)
p
(
q
|
u
)
Since the denominator
p
(
q
|
u
)
does not affect ranking, and
q
has no direct relation
with
u
,wehave
u
ʸ
Q
∝
p
(
z
|
u
)
p
(
q
|
z
)
∝
p
(
z
|
u
)
p
(
w
i
|
z
)
(4.13)
w
i
∈
q
where query
q
=
w
1
,...,
w
n
q
,
p
(
z
|
u
)
and
p
(
w
i
|
z
)
can be directly obtained from
w
. Note that we actually
user-topic distribution
ʩ
u
and topic-word distribution
ʦ
utilize the unigram language model in this chapter.
From the standard language model-based information retrieval in [
14
], image
model
v
D
can be directly represented as the topic distribution of the image visual
content
v
d
=
ʸ
v
d
1
,...,
v
dn
d
:
n
d
n
d
1
n
d
p
(
z
)
p
(
v
di
|
z
)
v
ʸ
D
p
(
z
|
v
d
)
=
p
(
z
|
v
di
)
=
(4.14)
n
d
p
(
v
di
)
i
=
1
i
=
1
where
n
d
indicates the number of visual descriptors in image
d
,
p
(
v
di
|
z
)
is the topic-
visual descriptor distribution,
p
(
z
)
is the topic prior distribution, and
p
(
v
di
)
is visual
descriptor prior distribution.
3
We consider noisy issue of user-generated annotations when representing image
model
w
D
. Besides aggregating each tag's topic-word distribution, we incorporate
the annotator authority as annotation confidence as weight for the corresponding tag
word. Formally, we have:
ʸ
n
d
p
(
z
)
p
(
z
|
u
w
di
)
p
(
w
di
|
z
)
w
ʸ
D
p
(
z
|
w
d
)
=
(4.15)
n
d
p
(
w
di
)
i
=
1
where
n
d
indicates the number of tag word in annotation of image
d
,
u
w
di
is the
annotator for tag
w
di
. Since user's dominant topic distribution can be viewed as
his/her expertise, we employ the annotator's topic distribution
p
to represent
the authority on topic
z
, which determines the trust we place on each tag. Now
(
z
|
u
w
di
)
3
We assume that visual descriptor prior and tag word prior all follow uniform distribution, i.e.,
p
1
|
V
|
1
|
W
|
(
v
di
)
=
,
p
(
w
di
)
=
.