Cheng et al. [6] do an analysis of more than 2.5 million of
Youtube videos, it all obtained from a crawler. The authors
evaluate some characteristics as category popularity and
number of views. Furthermore, they investigate the social
network of Youtube videos, created by related videos and
user generated content. This network has characteristics of
“small-world” and a high clustering coefficient, which means
that this behavior can be explored to improve the design of
caching and “peer-to-peer” strategies for video sharing.
The works of Acharya et al. [1] and Chesire et al. [7] have
focus on popularity studying. The first one made analysis
based on user access to videos streamed on the Web, and
it identified that the content popularity does not have the
Zipf distribution. The last one analyzes the server workload
of medias from a large company, ant it observed that the
popularity has a Zipf distribution. Both works have opposite
results, which can be explained by the different nature of
the evaluated content, and it demonstrates the impact of the
scenario of study.
Another research [4] presents a workload characterization
of a online video sharing system. The authors use a real
and representative workload to characterize access patterns
and to study the user navigation profiles of this system.
As results, they provide several statistical models to various
system characteristics, such as popularity of videos, users,
and tags, inter-request and inter-session time distributions,
etc. They show that a typical user session of online video
social networks remains about 40 minutes, corresponding
to a higher value to the session timeout of traditional Web
systems. Their analyses have novelties for online video
sharing systems and useful for synthetic workload generation
and to the project of new infra-structures of this kind of
service.
Addressing the recommendation problem, it is possible to
find lots of works with the purpose to present techniques
for recommendation systems. The Recommender Systems
Handbook [14] was published in 2011, and it contains a set
of papers that is embrace by five topics: techniques, application
and evaluation of recommendation systems; recommendation
systems interactions; recommendation systems and
communities; and advanced algorithms. This study addresses
subjects that compose the base of recommendations systems,
and it has been used as reference to application and development
of the method of recommendation that is used in
our research.
The work of Su and Khoshgoftaar [16] presents several
techniques of Collaborative Filtering (CF), that is one of the
most successful approach to build recommendation systems.
By describing its mainly advantages and disadvantages, the
author enumerate the most important techniques of CF:
memory-based, model-based and hybrid (combining the first
two).
One of the current challenge is the modeling of the
user behavior. Recommendation systems are based on pro-