Database Reference
In-Depth Information
make accurate predictions about the ranking of videos. To extend our model, one
can easily add further features or measurements such as geographical location,
social network interactions data, buffering ratio, rate of buffering events, session
join time, rendering quality, rate of bitrate switch,
etc.
Outputs.
The supervision
associated to each input video
x
isbasedonfour
possible ordered values which gives an indication for the final target ranking. In
our model,
y
Y∈{
0
,
1
,
2
,
3
}
, whose labels are
{
non-popular
,
popular
,
very popular
,
viral
respectively. It represents a natural ranking for Internet videos. Using
this ranking model, we intend to provide a measure of video demand, which is
closely related not only to the popularity, but also to the consumption of system
resources.
Finally, the learning-to-rank module finds a function
}
f
from Eq. (
1
) with
∀i, j, i
j, y
i
>y
j
the constraint of maintaining the prediction order:
=
then
f
>f
(
x
j
) explained in [
7
]. In that case, theoretical performance guaran-
tees are provided. Practically, the use of the mean square error (
(
x
i
)
y − f
(
x
))
2
instead of the indicator function
I
{.}
(which is hard to optimize because it is
non-differentiable) allows us to ensure a calibrated learning to rank algorithm.
A calibrated algorithm means there is a theoretical link between the approxima-
tion of the empirical risk, that is easier to optimize, and its non-differentiable
version [
7
,
9
,
31
,
35
].
3.3 Framework for Learning and Predicting, and Implementation
We implement our model using
ensemble methods
. According to Friedman
et al.
,
ensemble learning consists of a set of very popular supervised methods, that
are robust, simple to train and tune, and have a remarkable prediction per-
formance. Our implementation is based on Scikit-learn, a general-purpose
machine learning library [
26
].
We designed a simple framework to use our learning module, depicted in
Fig.
1
. Our framework has two phases: (i) learning and (ii) predicting. Each
phase has its own YouTube-like workload. Learning is a preliminary phase that
commonly runs oine in a batch mode, while the prediction can go online. In
this work, both phases are performed with data from simulations. In the learning
phase, we first generate the training dataset, described in Subsect.
5.4
. Then we
feed this training dataset to our learning model, represented here as module of
Fig. 1.
Framework for learning and predicting ranking of Internet videos.
Search WWH ::
Custom Search