Information Technology Reference
In-Depth Information
By this setting, if the searcher owns more #tag than the average #tag of his/her
influencers,
ˁ>
.
(
,
,
)
0
5, and
R
q
u
d
will contribute more the final rank list. Other-
wise, the influencer's preference
R
(
q
,
c
,
d
)
will be emphasized more.
4.5.3.1 Personalized Image Search
Since users' tagging activities indicate their personal relevance judgement, we
employ social annotations for personalized search evaluation. The main assump-
tion is that the images tagged by user
U
i
with tag
w
j
will be considered relevant if
U
i
issues
w
j
as a query. The randomly selected 100 users who tagged 50-100 images
constitute the test user set
U
test
. The overlapping 21 tags the 100 users used constitute
test query set
T
test
. In order to reduce the dependency between user annotation and
evaluation, for the training process, we remove the tagging data related to the test
queries.
Based on the proposed personalized image search framework, we consider the
following four settings:
•
w
D
ʸ
Basic
: basic personalized image search based on Eq. (
4.11
), which computes
without considering annotation confidence;
•
Basic
annotation confidence (
Basic_AC
): basic personalized image search lever-
aging annotator authority as annotation confidence in Eq. (
4.15
);
+
•
Social network with global influence (
Social_global
): social-based personalized
image search by considering global influence from contact users, which modified
Eq. (
4.16
)as:
)
R
1
−
ˁ
|
C
u
|
R
(
q
,
u
,
d
)
=
ˁ
R
(
q
,
u
,
d
)
+
p
(
Z
t
|
u
,
c
(
q
,
c
,
d
)
t
c
∈
C
u
We estimate the global influence by simply aggregating influences over topics
which is irrelevant to queries;
•
Social network with topic-sensitive influence (
Social_topic
): social-based per-
sonalized image search by considering topic-sensitive influences computed from
mmTIM.
We use mMAP as the evaluation metric, which is the mean of the mean average
precisions (MAP) for all test users. mMAP is defined as:
|
T
test
|
|
U
test
|
i
j
=
1
AP
(
i
,
j
)
|
T
test
|
|
U
test
|
=
1
mMAP
=
(4.19)
where AP
denotes the average precision value of the
j
th test query for the
i
th
user. The results are shown in Fig.
4.10
a. It is shown that
Basic_AC
outperforms
Basic
by 20%, which demonstrates the effectiveness of addressing the noisy annotation
issue by incorporating annotator authority. By further considering the influence from
(
i
,
j
)
