Information Technology Reference
In-Depth Information
that is well recognized to govern the behaviors of users and thus play important
role in personalized services. A probabilistic generative model is proposed to simul-
taneously address the above multimedia content analysis and user modeling tasks.
After that, we propose a theoretical framework for personalized image search. The
new framework benefits from the consideration of user preference and annotation
confidence in the modeling process. The derived social influence, multimedia topic
distribution, and user expertise are jointly modeled in the novel framework.
Social influence analysis has attracted extensive research interest, such as inves-
tigating the existence of influence [
1
,
9
], maximizing influence propagation [
5
,
12
],
and identifying the evolved influencer [
19
]. For social influence-based retrieval, the
basic premise behind is that the preferences of other users, who are the influencers to
the searcher, provide a good indication for the searcher's preference. The influence
link between influencers and searcher has two intrinsic characteristics: (1) The influ-
ence is continuous. This is easy to understand, as binary influence link (i.e., influencer
or not) is too coarse to model the relationship strength. Recently, continuous influ-
ence modeling has been addressed in [
31
,
32
]. (2) The influence is topic-sensitive.
Given a one-way influence network, the actual influencer is task-dependent.
In this chapter, we investigate the problem of topic-sensitive influencer mining in
social multimedia communities. In particular, we take Flickr, one of the most popular
photo sharingweb sites, as the social multimedia platform. In Flickr, users are allowed
to add others as their contacts, which is analogous to the influencers. Besides the
explicit one-way influence link between users, the uploaded images and associated
tags are leveraged. We propose a probabilistic generative model to infer the output,
which inverses the generation process of image content as well as the associated
tags. With users as nodes, user-influencer social links as edges, the output includes
(1) the learned topic space, (2) the topic distribution of each node, and (3) the topic-
sensitive edge strength. After mining the topic-sensitive influencer in the contact
network, we apply it to the application of personalized search. Following the risk
minimization-based information retrieval scheme, we present a general framework
for personalized image search. The new framework benefits from the consideration
of influencer preference and annotation confidence in the modeling process. Topic-
sensitive influence and user expertise can be readily integrated in the expansion of
the Language Models (LM) for queries and image documents. Part of the work in
this chapter has been published in [
26
,
27
].
4.2 Related Work
Extensive efforts have been focusing on personalized search these days. The resources
being leveraged include explicit user profile, relevance feedback, click-through data,
context information, social annotation, and social network. Reference [
6
] conducted
an early work by re-ranking search results according to the cosine distance between
eachURL and explicit user interest profiles. Reference [
13
] utilized the search context
