User-Perceptive Multimedia Content Analysis - User-centric Social Multimedia Computing - page 14

utilized, which assumes that people tend to catch others emotions as a consequence of
facial, vocal, and postural feedback. In [ 9 ], the directed social relations, i.e., “contact”
in Flickr, is exploited to estimate the photo popularity and influence at topic-level.
Hypergraph is constructed to represent user, image and the heterogeneous relations
between users and images. Some studies integrate the metadata, usage data and
interactions between users towards social multimedia analysis. In [ 10 ], the authors
proposed to exploit the heterogeneous information like users' tagging behaviors,
social networks, tag semantics and item profiles to alleviate the cold start problem in
recommender system. In [ 37 ], a hybrid solution is presented to identify the geographic
location of web videos. Various available sources of information, e.g., user's upload
history, the social network, video tagging, are exploited with a divide & conquer
strategy.
2.2.2 Social Image Tag Refinement
The literatures [ 22 , 47 ] provide good surveys for the research work on image tag
refinement. Along the three basic elements in photo tagging behaviors, i.e., image,
tag, and user, we characterize the related work according to the elements they lever-
aged.
As a pioneer work, Jin et al. [ 16 ] employed WordNet to estimate the semantic
correlations among the annotated tags and remove weakly correlated ones. The work
of [ 39 ] performs belief propagation among tags within the random walk with restart
framework to refine the imprecise original annotations. In [ 42 ], Xu et al. proposed
to jointly model the tag similarity and tag relevance and perform tag refinement
from the topic modeling view. These work is typically based on the tag - tag analysis.
In [ 24 ], the authors explicitly considered the tag-image and tag-tag relations and
proposed a dual cross-media relevance model for image annotation. Liu et al. [ 21 ]
proposed to rank the image tags according to their relevance w.r.t. the associated
images by modeling tag similarity and image similarity. In [ 20 ], the improved tag
assignments are learnt by maximizing the consistency between visual similarity and
semantic similarity while minimizing the deviation from initially user-provided tags.
An interesting work is done by Xie et al. [ 41 ], in which several important issues in
building an end-to-end image tagging application are addressed, including tagging
vocabulary design, taxonomy-based tag refinement, classifier score calibration for tag
ranking, and selection of valuable tags. Recently, Liu et al. [ 23 ] proposed a multiedge
graph based unified framework to solve the image annotation, tag-to-region and tag
refinement problem. Tag - tag , image - image and image - tag relationships are explored
in these work.
The most related work to this chapter is [ 19 , 47 ], which solves the tag refine-
ment problem through low-rank matrix approximation. Zhu et al. [ 47 ] considered
the tagging characteristics from the view of low-rank, error sparsity, content consis-
tency and tag correlation. In [ 19 ], a factor analysis model is proposed and the tag
refinement problem is cast as estimating the image-tag correlations.While these work