TATAR Alexandru Florin
Team : NPA
Departure date : 10/15/2014
Supervision : Serge FDIDA
Predicting User-Centric Behavior: Content Popularity and Mobility
Understanding user behavior is fundamental in the design of efficient communication systems. Unveiling the complex online and real-life interactions among users, deciphering online activity, or understanding user mobility patterns — all forms of user activity — have a direct impact on the performance of the network. But observing user behavior is not sufficient. To transform information in valuable knowledge, one needs however to make a step forward and go beyond observing and explaining the past to building models that will predict future behavior. In this thesis, we focus on the case of users consuming content on the move, especially when connectivity is poor or intermittent. We consider both traditional infrastructure-based communications and opportunistic device-to-device transfers between neighboring users. We offer new perspectives of how to use additional information about user behavior in the design of more efficient solutions for mobile opportunistic communications. In particular, we put forward the case that the collective user behavior, both in terms of content consumption and contacts between mobile users, can be used to build dynamic data replication strategies.
We first investigate content consumption patterns. Our contribution in this area is two-fold. First, we analyze a large news data set published on 20minutes.fr, a popular daily newspaper in France. We survey the different prediction algorithms proposed in the literature and compare the ability of two of these methods to predict the popularity of online news articles. We observe that a linear model on a logarithmic scale is an effective solution to predict the popularity of online news. Furthermore, in the context of automatic online news ranking we observe that this method is also an effective solution to correctly rank items based on their future popularity with a performance that can evenly match more customized learning-to-rank algorithms. We study then the practical impact of using a model that can predict content popularity in a mobile opportunistic scenario. We place this in the context of mobile data offloading where the goal is to proactively seed content during idle periods to reduce data traffic during the peak periods. We show that the ability to actually predict future user demand can improve the benefit of proactive seeding for a mobile opportunistic data offloading solution compared to traditional methods that consider a stable evolution of content popularity.
In a mobile scenario, users who share common interest in a content and are within physical proximity, can establish device-to-device connections and retrieve content directly from their neighbors. We study the predictability of human contacts. User mobility, represented as a highly dynamic system, is not completely random and patterns can be learned after studying user movement for a certain period of time. But contacts between mobile users are a scarce resource, as some users will often come close to each other but never in direct contact. We thus extend the prediction task to the multi-hop contact case — predict if mobile users will find themselves at a distance of at most k-hops from one another. By analyzing three real-life contact traces we observe that, in a mobile scenario characterized by frequent disconnections, one can obtain better performance when predicting that nodes will find themselves at a greater distance from one another compared to the direct contact case. To assess the impact of these findings in a real-life scenario, we propose a simulation experiment in which, by combining mobile opportunistic communications with k-contact prediction, one can reduce the amount of traffic used in the communication of mobile nodes with the cellular infrastructure.
: 07/09/2014 - 14h00 - Site Jussieu - Salle Jean-Louis Laurière - 25-26/101Jury members
Walid Dabbous, INRIA Sophia Antipolis
Andrea Passarella, IIT-CNR
Martin May, Technicolor
Anne-Marie Kermarrec, INRIA Rennes
Sebastien Tixeuil, UPMC Sorbonne Universités
Marcelo Dias De Amorim, CNRS and UPMC Sorbonne Universités
Serge Fdida, UPMC Sorbonne Universités