Information Technology Reference
In-Depth Information
we have introduced the basic personalized image search framework which is based
on risk minimization. As discussed at the beginning of this section, one important
advantage of language model is its expansibility. We have successfully encoded
annotation confidence to tackle noisy user-generated social tags by Eq. (
4.15
). In the
following, we will discuss how to incorporate the topic-sensitive influence analysis
between users into the proposed framework.
Based on the above formulation, we can compute a personalized risk value for
each individual image when searching by
u
as
R
. In the context of influence
network, searcher's influencer will impact the personalized rank by modifying the
risk as follows:
(
q
,
u
,
d
)
)
R
1
−
ˁ
|
C
u
|
R
(
q
,
u
,
d
)
=
ˁ
R
(
q
,
u
,
d
)
+
p
(
Z
t
|
u
,
c
)
p
(
Z
t
|
q
(
q
,
c
,
d
)
t
∈
C
u
c
(4.16)
where
R
is weight term balanc-
ing contributions from the searcher and his/her influencers,
p
(
q
,
c
,
d
)
is the risk value when searching by user
c
,
ˁ
is the topic-
sensitive influence strength between user
u
and
c
on topic
Z
t
. The intuition here is
that, if one influencer has high influence to the searcher on the query-concerning topic
(
Z
t
|
u
,
c
)
t
)
should be emphasized and attached larger weight. Recall the toy example in Fig.
4.1
,
since word “Tahiti” closely relates to the
travel
topic, when Bob searches “Tahiti”,
Tom's suggestion will be highly valued and contribute most to the final decision.
)
, his/her preference to each candidate images
R
p
(
Z
t
|
u
,
c
)
p
(
Z
t
|
q
(
q
,
c
,
d
4.5 Performance Evaluation
4.5.1 Dataset
The data are collected from a large-scale Flickr dataset NUS-WIDE [
7
], which con-
sists of 269,648 images. We crawled the image's uploader information and obtained
50,120 unique users. Since the focus in this chapter is social influence analysis from
user perspective, we remove the users uploading less than 15 images and conduct
experiments on the remaining users and their images. For each user, we collected
his/her annotation set and contact list. If there exists contact relation between two
users, we refer to it as one contact edge. The statistic of resulted dataset is shown in
Table
4.1
.
In this work, we choose to represent the image visual content by region-level
maximally stable extremal region (MSER) feature. Compared with key-point based
descriptor, MSER regions indicate local homogeneous parts in objects and shown
Table 4.1
The statistics of the collected Flickr data
#user
#contact edge
#image
#tag token
#unique tag
3,372
53,952
124,099
623,254
30,108
