Information Technology Reference
In-Depth Information
tags. In [
13
], the authors studied the problem of constructing tag hierarchies from
social tagging data, and investigated into the usefulness of such tag hierarchies in
supporting efficient navigation from an information retrieval perspective. In [
29
], the
issues of sparse, shallow, ambiguous, noisy, and inconsistent are considered when
exploiting the structured metadata to folksonomy learning. A relational clustering
solution is proposed. The third line is on exploiting the geographical metadata, e.g.,
geographical tag or check-in record. Ye et al. [
44
] develops a semantic annotation
technique for location-based social networks to automatically annotate all places
with category tags. In the Livehoods Project [
6
] exploits the check-in records col-
lected from a location-based social network to study the structure and composition of
a city, which requires long hours of observation and interviews in traditional means.
Another interesting line is on interpreting and modeling the tag creation process from
a generative perspective. One inspiring work is [
26
], where a new probabilistic gen-
erative model is proposed to simulate the generation process of social annotations
from both resource topics and user perspectives.
Another important type of user interaction is user usage data recorded during
social multimedia activities. Users unintentionally embed their understanding of the
multimedia content in their interaction. Related works are reviewed along three lines.
(1) Exploiting the browsing behaviors. An early work is conducted to utilize video
browsing log, e.g., pause, forward, to understand the video semantic structure, and
with applications in video browsing and summarization [
45
]. In [
3
], the authors
investigated into the impact of previous views to the future popularity of YouTube
videos, and found a “rich-get-richer” phenomenon. Inspired by the phenomenon,
[
28
] presents a solution to employing the early view patterns to predict the future
popularity, with potential applications in targeted advertising, effective search and
recommendation services. (2) Exploiting social endorsements. Many social multi-
media networking platform allows users to endorse entities that they find appealing.
Reference [
17
] exploits the typical social endorsement activities, “favoriting” photos
in Flickr, to extract relevant and descriptive entity semantics. In [
15
], a system called
“LinkMiner” is developed to employ Facebook “Like” to understand use interest and
estimate representativeness and influence of objects. (3) Exploiting comments. Refer-
ence [
7
] conducts a pioneer work on examining the motivations that users participate
into YouTube video commenting conversations. In [
30
], the authors introduced an
interesting work to exploit comments to analyze the cross-media similarity between
textual and video items. Instead of focusing on cross-modal association analysis, the
associated comments are employed as bridges for cross-media analysis and retrieval.
Moreover, comments have also been exploited in other multimedia content analysis
tasks, including inferring video semantics [
8
,
11
], estimating the mood of music or
video [
36
,
43
], predicting item popularity [
12
], and so on.
Besides the users' direct interaction with social multimedia objects, very recently,
the interactions between users are also exploited to address the social multimedia
content analysis tasks. The social interactions between users, i.e., social relations,
can be categorized into undirected and directed relations. In [
14
], the authors has
explored the utilization of undirected social relations to facilitate sentiment analysis in
the context of microblogging. A social science theory called “emotion contagion” is
