Links streams for the modelling of interactions over time and application to the analysis of IP traffic
Interactions are everywhere: in the contexts of face-to-face contacts, emails, phone calls, IP traffic, online purchases, running code, and many others. Interactions may be directed, weighted, enriched with supplementary information, yet the baseline remains: in all cases, an interaction means that two entities u and v interact together from time b to time e: for instance, two individuals u and v meet from time b to time e, two machines on a network start an IP session from time b to time e, two persons u and v phone each other from time b to time e, and so on. In this thesis, we explore a new approach consisting in modelling interactions directly as link streams, i.e. series of quadruplets ( b, e, u, v ) meaning that u and v interacted from time b to time e, and we develop the basis of the corresponding formalism. In order to guide and assess this fundamental work, we focus on the analysis of IP traffic. It is particularly important to us that we make both fundamental and applied progress: application cases should feed our theoretical thoughts, and formal tools are designed to have meaning on application cases in the most general way. We apply our framework to the analysis of IP traffic, with the aim of assessing the relevance of link streams for describing IP traffic as well as finding events inside the traffic. We devise a method to identify events at different scales, and apply it to a trace of traffic from the MAWI dataset.
Defence : 09/28/2016 - 14h - Site Jussieu 25-26/105 Jury members : Pierre Borgnat, CNRS/ENS Lyon [Rapporteur]
Olivier Festor, Inria/Telecom Nancy [Rapporteur]
Arnaud Casteigts, Université de Bordeaux
Marcelo Dias de Amorim, CNRS/UPMC
Matthieu Latapy, CNRS/UPMC
Clémence Magnien, CNRS/UPMC
Philippe Owezarski, LAAS
Véronique Serfaty, DGA.