Information Technology Reference
In-Depth Information
information collected from users' feedback to refine the raw query for second search.
Reference [
23
] utilized user's click-through history to construct user preference
vector and extended Topic-Sensitive PageRank for personalized search. Reference
[
15
] introduced a local search system by considering context information (e.g.,
time, location, weather) as well as user query log for search result personalization.
Reference [
16
] demonstrated that user-generated metadata indicates user's interest
and can be used to personalize information. They extracted latent topics from tagging
data and used the latent user interest profile for personalized image search. In similar
spirits, [
18
] utilized user activities of joining interest group to extract latent inter-
est profile and re-rank the returned images by combining latent interest-based user
preference and query relevance. Recently, an interesting work is performed in [
28
],
where the authors proposed to extract user-specific topic spaces by expanding the
raw annotation set for each user. Queries are mapped to the derived user-specific
topics spaces and the personalized relevance score over images are calculated.
Several approaches have directly or indirectly employed user's social relations for
personalized search. Reference [
2
] proposed a PageRank-like algorithm to exploit
the utilization of social relations. They assumed that a document receives an extra
“friendship” score when tagged by the searcher's friends. Reference [
4
] explicitly
defined familiar and similar scores to model relations between users. With familiar
score estimated from explicit social link and similar score calculated from collabora-
tive activities, personalization is conducted according to the searcher's relationship
strength with users in the social network. Similarly, in [
11
], a multilevel actor simi-
larity method is proposed to first measure the social affinity between users. The per-
sonalized ranking score is then calculated by combining the query-video relevance
and the affinity score between the searcher and the video's owner. Reference [
21
]
leveraged the output of community discovery and proposed a community-oriented
re-ranking scheme to aggregate query-video relevance and user-community-video
preference.
Our work differentiates from existing social network-based approaches in that,
we exploit the topic-sensitive social relations between users and integrate query-
dependent social strength into personalized ranking score calculation. Moreover,
for the first time, we present a theoretical personalized image search framework by
jointly modeling social influence preference and annotation confidence.
4.3 Topic-Level User and Multimedia Content Modeling
This section introduces the topic-sensitive influencer mining problem. First, we for-
mally define the problem:
Definition 1
(
Topic-sensitive Influencer Mining
) Given a collection of Flickr users
U
, each user
u
∈
U
corresponds to a three-dimensional tuple
[
C
u
, D
u
, T
u
]
.The
w
T
×|
W
|
goal of topic-sensitive influencer mining is to learn (1) Topic space
ʦ
ↂ R
