Dans ce chapitre, nous présentons une méthode gloutonne pour optimiser la modularité d'un graphe. Cette méthode de partionnement permet de traiter avec une excellente précision des systèmes de taille inégalée, allant jusqu'à plusieurs milliards de liens. Notre algorithme a de surcroît l'avantage de ne pas être limité à l'optimisation de la modularité puisqu'il peut être généralisé à d'autres fonctions de qualité, et de découvrir des communautés à différentes échelles. Les performances de l'algorithme sont évaluées sur des graphes artificiels pour lesquels la structure communautaire est connue, ainsi que sur des graphes de terrain réels.

Detecting events such as major routing changes or congestions in the dynamics of the internet topology is an important but challenging task. We explore here a {\em top-down} approach based on a notion of statistically significant events. It consists in identifying statistics which exhibit a homogeneous distribution with outliers, which correspond to events. We apply this approach to ego-centerd measurements of the internet topology (views obtained from a single monitor) and show that it succeeds in detecting meaningful events. Finally, we give some hints for the interpretation of such events in terms of network events.

Complex networks can often be divided in dense sub-networks called communities. We study, using a partition edit distance, how three community detection algorithms transform their outputs if the input network is sligthly modified. The instabilities appear to be important and we propose a modification of one algorithm to stabilize it and to allow the tracking of the communities in a dynamic network. This modification has one parameter which is a tradeoff between stability and quality. The resulting algorithm appears to be very effective. We finnaly use it on a dynamic network of blogs.

Most complex networks can be divided into dense sub-graphs called communities. These communities may also be divided recursively producing a hierarchical structure of communities, summarized in a tree named dendrogram. In this article we analyze this structure extracted from several complex networks. First we study the shape of the tree and, in particular, communities imbrications. Then we show that an excessive decomposition of communities can result in meaningless communities. We propose a couple of approaches to solve this problem.

Many contributions rely on the degree distribution of the Internet topology. However, current knowledge of this property is based on biased and erroneous measurements and is subject to much debate. Recently, a new approach, referred to as the Neighborhood Flooding method, was proposed to avoid issues raised by classical measurements. It aims at measuring the neighborhood of Internet core routers by sending traceroute probes from many monitors distributed in the Internet towards a given target router. In this paper, we investigate the accuracy of this method with simulations. Our results show that Neighborhood Flooding is free from the bias highlighted in the classical approach and is able to observe properly the exact degree of a vast majority of nodes in the core of the network. We show how the quality of the estimation depends on the number of monitors used and we carefully examine the influence of parameters of the simulations on our results. We also point out some limitations of the Neighborhood Flooding method and discuss their impact on the observed distribution.

Maps of the internet topology are generally obtained by measuring the routes from a given set of sources to a given set of destinations (with tools such as traceroute). It has been shown that this approach misses some links and nodes. Worse, in some cases it can induce a bias in the obtained data, i.e. the properties of the obtained maps are significantly different from those of the real topology. In order to reduce this bias, the general approach consists in increasing the number of sources. Some works have studied the relevance of this approach. Most of them have used theoretical results, or simulations on network models. Some papers have used real data obtained from actual measurement procedures to evaluate the importance of the number of sources and destinations, but no work to our knowledge has studied extensively the importance of the choice of sources or destinations. Here, we use real data from internet topology measurements to study this question: by comparing partial measurements to our complete data, we can evaluate the impact of adding sources or destinations on the observed properties.

We show that the number of sources and destinations used plays a role in the observed properties, but that their choice, and not only their number, also has a strong influence on the observations. We then study common statistics used to describe the internet topology, and show that they behave differently: some can be trusted once the number of sources and destinations are not too small, while others are difficult to evaluate.