Abstract |
The current traffic trends and predictions foresee many challenges for future mobile
networks, which will need to efficiently serve traffic volumes orders of magnitude larger
than those experienced today. Network-aware Recommendations has been recently
proposed as a paradigm that enables mobile networks to keep up with the increasing data
demand, by jointly designing communication networks and recommendation systems
(RSs).
Network-aware recommendations is a relatively new research area having many
underexplored topics and open research problems, some of them having been discussed
in this article. Network-aware recommendations (i) are based on the fact that
recommendation systems drive a significant fraction of the demand for content in the
Internet (e.g., more than 50% of user requests at YouTube come from its
recommendations and the respective percentage for Netflix is 80%), and (ii) steer
recommendations towards content that can be delivered efficiently through the networks
(e.g., locally stored at a cache in the mobile edge or exploiting coded transmissions). Most
of the related work of the field applies in theoretical scenarios, without making realistic
evaluations, including real-world setups and real-user ratings. In this work we are the first
to experimentally evaluate (through measurements in a real service and experiments
with users) the performance of proposed network-aware recommendation approaches
and investigate their feasibility and benefits from different points of view (such as
network performance, user experience, etc.).
In detail we:
1. leverage public information provided by the YouTube API Service and conduct
realistic simulations to evaluate the potential gains from the network perspective
2. implement and use an experimental testbed to interact with real users and
demonstrate the benefits for the content providers’ and end users’ perspective,
respectively
3. build statistical models to derive the user experience as a function of QoS and
user interest and to provide useful insights for the design of network-aware RSs
We believe that our study is an important first step towards network-aware
recommendations, by providing experimental and analytic evidence for their feasibility
and benefits and by discussing the interplay between networking and content
recommendations.
|