Information Technology Reference
In-Depth Information
1.2 Challenges and Progresses
Social multimedia data exhibit unique characteristics and pose great challenges to
social multimedia computing and potential applications.
From the perspective of generation, social multimedia is noisy and diverse.
The
user-generated mechanism gives rise to the issues of low quality as well as huge
quantity. Users with various background use social media to record their daily life,
resulting in subjective social multimedia data and featuring a wide array of attributes
like resources, appearance, and degree of diffusion. Basic social multimedia com-
puting tasks, e.g., social image classification, suffer from the diversity characteristic,
since images describing the same concept may show very different appearance. The
notorious intravariance issue significantly challenges the traditional classification
algorithms. Moreover, the huge quantity makes it very difficult to locate the desired
multimedia content. How to understand and exploit the noisy, diverse, and large-scale
social multimedia data becomes critical in social multimedia computing.
From the perspective of distribution, social multimedia is heterogeneous.
Besides
textual data, social media involves heavily with multimedia data, such as photo shar-
ing websites Flickr and Pinterest, video sharing websites YouTube and Hulu, and
music sharing websites Last.fm and FreeSound, etc. It is very common for multi-
modal data, e.g., text, image, and video, to exist simultaneously in one social media
platform. Moreover, emerging social media services have also given birth to social
multimedia data with novel modality, such as location data in check-in services.
Data with different modalities challenges integrated social multimedia computing.
Take user modeling as example. To understand user preference from his/her online
activities, desired solutions need to model the heterogeneous user data (e.g., regis-
tration profile, browsing history, images and videos uploaded, and comments and
annotations) in a principled way.
From the perspective of interaction, social multimedia is interconnected.
We refer
to it as the “collective” effect that social media data does not exist independently
but interacts with each other. It is the core characteristic that social multimedia data
differentiates from traditional multimedia data. The collective effect is either explicit
or implicit. For example, the interaction of observed user-user relationship via their
connecting behaviors is explicit, while the interaction by collaboratively annotating
the same image or choosing the same tags is implicit. The collective effect among
social multimedia data violates the independently and identically distributed (i.i.d.)
assumption in most statistical machine-learning algorithms. Due to this collective
effect, social multimedia data has remarkable social attributes, where both content
and collective information need to be considered for effective social multimedia
computing.
To address the above challenges, efforts have been taken in the research lines
of (1) social multimedia content analysis, (2) user understanding, and (3) collective
search and recommendation. In the following, we highlight some typical work to
briefly introduce the progresses.
