Information Technology Reference
In-Depth Information
tags are considered in a conservative filtering strategy by exploiting the tag correla-
tion on context and semantics. Besides, the multiple intrarelations are employed as
the smoothness constraints and then the factors inference is cast as a regularized ten-
sor factorization problem. Finally, based on the learnt factor representations, which
encode the compact users, images, and tags representation over their latent subspaces,
tag refinement is performed by computing the cross-space
image
-
tag
associations.
Most of the work in this chapter has been published in [
33
,
34
].
2.2 Related Work
2.2.1 Multimedia Content Analysis
The idea of exploiting user perception has been realized in many social multimedia
content analysis tasks. In this subsection, as summarized in Table
2.1
, we will review
related work based on the exploited user interactions.
The first type of user interaction is the metadata associated with social multimedia
content, e.g., descriptions, tags. Such metadata provides a natural context for mul-
timedia content analysis, which helps reducing the semantic gap between low-level
multimodal features and the high-level semantics. Since the user-contributed tags are
generally noisy, ambiguous, incomplete or subjective, one critical research topic in
exploiting themetadata is tag processing. Typical tag processing problems include tag
ranking [
21
], tag refinement [
47
], tag-to-region [
25
], etc.
2
After tag processing, the
second research line is devoted to semantic ontology construction from the processed
Table 2.1
The categorization
of the work on
user-perceptive multimedia
content analysis according to
the exploited user interactions
Exploited user
Related work
interaction
Metadata
Tag processing [
21
,
25
,
47
]
Semantic ontology construction [
13
,
29
]
Multimedia check-in mining [
6
,
44
]
Tag generation modeling [
26
]
Usage data
Browsing behavior [
3
,
28
,
45
]
Social endorsements [
15
,
17
]
Comments [
7
,
8
,
11
,
12
,
30
,
36
,
43
]
User-user
Undirected relation [
14
]
interaction
Directed relation [
9
]
Hybrid [
10
,
37
]
2
Note in most tag processing work, while tag is contributed by users, user factor is not explicitly
considered. We will discuss the difference between our work in this chapter and the existing tag
process work in next subsection.
