Massoulie Laurent (Thomson)

Content popularity and user profiling for content placement 

in peer-to-peer VoD systems


Based on joint work with Peter Marbach and Dan-Cristian Tomozei

In this talk we consider the problem of pre-loading video content at users of a peer-to-peer system, so as to meet content demands with minimal access to infrastructure servers. We propose to leverage content popularity distribution as well as user profiling to obtain maximal benefits.

We will discuss optimisation formulations, and resulting replication strategies that exploit expected demand and its volatility. We will also discuss ongoing work on user profiling, in which we advocate the use of particular spectral clustering techniques to identify profiles of users. Corresponding theoretical guarantees as well as empirical validation on the NetFlix dataset will be presented.

Download presentation files