Johan Mazel

Mercredi 8 juin 2016 à 11h, salle 24-25/405

Internet plays a central role in our lives. However, it is an extremely diverse and complex system. Ranging from non-malicious unexpected events such as flash-crowds and failures, to network attacks such as denials-of-service and network scans, network traffic anomalies can have serious detrimental effects on the performance and integrity of the network. Anomaly detection is thus paramount in order to guarantee users’ access to Internet resources. In this talk, we will address recent advances in network traffic anomaly detection and classification, that leverage graph analysis techniques, machine learning techniques and big data.

Argyris Kalogeratos

Lundi 13 juin 2016 à 11h, salle 26-00/332

Slides

In many real-life situations, it is critical to dynamically suppress or remove an undesired diffusion process (viruses, information, behaviors, etc.). The talk will present a framework for Dynamic Resource Allocation (DRA) assuming a continuous-time SIS epidemic model, and that a budget of treatment resources of limited efficiency are at the disposal of authorities. Special emphasis will be given on the macro- and microscopic (or local) properties of the network structure for the problem and two strategies will be presented that fall in this framework: a) a simple yet effective greedy approach, and b) a more sophisticated one that uses a precomputed priority plan of how the healing strategy should proceed on a specific network. Additionally, extensions in competitive scenarios will be discussed.

Marthe Bonamy

Vendredi 3 juin 2016 à 11h, salle 24-25/405

Slides

Given a colouring of a graph, a Kempe change is the operation of picking a maximal bichromatic subgraph and switching the two colours in it. Two colourings are Kempe equivalent if they can be obtained from each other through a series of Kempe changes. Kempe changes were first introduced in a failed attempt to prove the Four Colour Theorem, but they proved to be a powerful tool for other colouring problems. They are also relevant for more applied questions, most notably in theoretical physics. Consider a graph with no vertex of degree more than some integer D. In 2007, Mohar conjectured that all its D-colourings are Kempe-equivalent. Feghali, Johnson and Paulusma proved in 2015 that this is true for D=3, with the exception of one single graph which disproves the conjecture in its generality. We settle the remaining cases by proving the conjecture holds for every integer D at least 4. This is a joint work with Nicolas Bousquet (LIRIS, Ecole Centrale Lyon, France), Carl Feghali (Durham University, UK) and Matthew Johnson (Durham University, UK).

Sergey Kirgizov

Lundi 04 julliet 2016 à 11h, salle 24-25/405

Slides

At first, we say that a ego-community structure is a probability measure defined on the set of network nodes. Any subset of nodes may engender its own ego-community structure around. Many community detection algorithms can be modified to yield a result of this type, for instance, the personalized pagerank. We also recall that community detection algorithms (including personalized pagerank) can be viewed from different perspectives: random walks, convergence of markov chain, spectral clustering, optimization, mincut(s), discrete cheeger inequality(ies), etc. Next, we present a continuous version of Viard-Latapy-Magnien link streams, that we call « temporal density ». Classical kernel density estimation is used to move from discrete link streams towards their continuous counterparts. Using matrix perturbation theory we can prove that ego-community structure changes smoothly when the network evolves smoothly. This is very important, for example, for visualization purposes. Combining the temporal density and personalized pagerank methods, we are able to visualize and study the evolution of the ego-community structures of complex networks with a large number of temporal links in order to extract interacting information. For example, we can detect events, trace the evolution of (ego-)community structure, etc. We illustrate and validate our approach using « Primary school temporal network data » provided by sociopatterns.org, and we show how the temporal density can be applied to the study of very large datasets, such as a collection of tweets written by European Parliament candidates during European Parliament election in 2014.

Jean Creusefond

Vendredi 13 mai 2016 à 11h, salle 24-25/405

Slides

Cet exposé sera en deux parties relativement indépendantes : l’évaluation des structures communautaires par le biais des vérités de terrain et l’analyse des l’appartenance communautaire des motifs dans les flux de liens L’évaluation de structures communautaires de manière théorique est très délicate : de multiples propriétés structurelles sont considérées comme importantes, par conséquent considérer une structure comme meilleure qu’une autre implique des choix arbitraires sur ces préférences, matérialisé par le choix d’une fonction de qualité ou de benchmarks. Afin d’éviter ces problèmes, beaucoup de chercheurs évaluent maintenant leurs résultats par comparaison avec des structures communautaires extraites en même temps que des jeux de données, en argumentant que la proximité entre leurs résultats et la vérité de terrain est une preuve significative de pertinence. Dans cette partie, je vais discuter d’une méthodologie permettant de concilier les deux approches et d’identifier quelles vérités de terrain favorisent quelles fonctions de qualité. Je soulignerai notamment le choix de la fonction de comparaison de partitions, souvent considéré comme anodin, mais changeant en fait radicalement les résultats. Pour référence, le programme développé (incluant un grand nombre d’algorithmes de détection de communautés et de fonctions de qualité) est entièrement disponible à l’adresse suivante : https://codacom.greyc.fr/ En seconde partie, je discuterai de travaux en cours d’analyse de flots de liens : des graphe dont chaque arc est étiqueté par un temps et où les multiarcs sont possibles. Les flots de liens qui nous intéressent ici représentent des réseaux de communication, c’est-à-dire que chaque arc représente une interaction orientée entre deux utilisateurs. Fréquemment, les algorithmes de détection de communautés qui tentent de les analyser agglomèrent le réseau de communication sur des fenêtres temporelles, où des méthodes traditionnelles (ou adaptées) peuvent êtres appliquées. Dans ce cas, une information est perdue : la causalité entre les liens. Par exemple, si un ensemble de personnes ont systématiquement la même structure de communication (ex : quand « A » interagit avec « B », celui-ci intéragit ensuite systématiquement avec « C » et « D »), peut-on en déduire la structure communautaire associée? Afin d’évaluer l’impact de cette information, je me suis intéressé aux motifs : des chaînes de communication dont la causalité semble probable (la première interaction a probablement entraîné la suivante, etc.). Le lien entre ces motifs et la structure communautaire reste donc à analyser, et je présenterai les outils mis au point à ce dessein ainsi que quelques résultats préliminaires.

Fragkiskos Malliaros

Vendredi 01 avril 2016 à 11h, salle 24-25/405

Slides

Networks have become ubiquitous as data from diverse disciplines can naturally be mapped to graph structures. The problem of extracting meaningful information from large scale graph data in an efficient and effective way has become crucial and challenging with several important applications and towards this end, graph mining and analysis methods constitute prominent tools. In this talk, I will present part of my work that builts upon computationally efficient graph mining methods in order to: (i) design models for analyzing the structure and dynamics of real-world networks towards unraveling properties that can further be used in practical applications; (ii) develop algorithmic tools for large-scale analytics on data with inherent (e.g., social networks) or without inherent (e.g., text) graph structure. Our approaches rely on the concepts of graph degeneracy and core decomposition in graphs. In particular, for the former point I will show how to model the engagement dynamics of large social networks and how to assess their vulnerability with respect to user departures from the network. In both cases, by unraveling the dynamics of real social networks, regularities and patterns about their structure and formation can be identified; such knowledge can further be used in various applications including churn prediction and anomaly detection. For the latter, I will present a core decomposition-based approach for locating influential nodes in complex networks, with direct applications to epidemic control and viral marketing.

Samuel Martin

Vendredi 18 mars 2016 à 11h, salle 26-00/332

Slides

Opinion evolution and judgment revision are mediated through social influence. In this talk, I will present a study based on a crowdsourced in vitro experiment. The study shows how a consensus model can be used to predict opinion evolution in online collective behaviour. The model is parametrized by the influenceability of each individuals, a factor representing to what extent individuals incorporate external judgments. Judgment revision includes unpredictable variations which limit the potential for prediction. The study also serves to measure this level of unpredictability via a specific control experiment. More than two thirds of the prediction errors are found to occur due to unpredictability of the human judgment revision process rather than to model imperfection.

Thomas Louail

Vendredi 19 février 2016 à 11h, salle 24-25/405

Slides

The increasing availability of individual geographic footprints has generated a broad scientific interest for human mobility networks, at various scales and in different geographical contexts. In this talk I will present some recent results related to urban mobility. I will first present a method we developed to extract an expressive and coarse-grained signature from a large, weighted and directed network. I will then discuss the results we obtained when we applied this method to mobility networks extracted from mobile phone data in 31 Spanish cities, in order to compare the structure of journey-to-work commuting in these cities. The method distinguishes different types of links/flows, and clearly highlights a clear relation between city size and the importance of these types of . In the second part of my talk, I will focus on the shopping trips networks extracted from credit card transactions, performed by hundreds of thousands of anonymized individuals in the two largest Spanish cities. Starting from the bipartite networks that link individuals and businesses of the city, I will show that it is possible to evenly distribute business income among neighbourhoods — and then mitigate spatial inequality — by reassigning only a very small fraction of each individual’s shopping trips. This spatial and bottom-up approach of wealth redistribution could be easily implemented in mobile apps, that would assist individuals in slightly reshaping their mobility routines. Our results hence illustrate the social benefits individuals are entitled to expect from the analysis of the data they daily produce.

Patricia Conde Céspedes

Vendredi 22 janvier 2016 à 11h, salle 26-00/332

Slides

Although the notion of community does not have a unanimous accepted definition, it is often related to a set of strongly interconnected nodes. Indeed, the density of links plays an important role because it measures the strength of the relationships in the community. The need of these well connected and dense communities has led to the notion of consensus community. An consensus community, is a group of nodes where each member is connected to more than a proportion of the other nodes. An consensus community is maximal if and only if adding a new node to the set breaks the rule. Consequently, an consensus community has a density greater than . Existing methods for mining consensus communities generally assume that the network is entirely known and they try to detect all such consensus communities. Detecting the local community of specific nodes is very important for applications dealing with huge networks, when iterating through all nodes would be impractical. In this paper, we propose an efficient algorithm, called RANK-NUM-NEIGHS (RNN), based on local optimizations to approximate the maximal consensus local community of a given node. The proposed method is evaluated experimentally on real and artificial complex networks in terms of quality, execution time and stability. We also provide an upper bound on the optimal solution. The experiments show that It provides better results than the existing methods.

Oana Bllu

Vendredi 27 novembre 2015 à 11h, salle 24-25/405

Slides

Finding dense subgraphs in large graphs is a key primitive in a variety of real-world application domains, encompassing social network analytics, event detection, biology, and finance. In most such applications, one typically aims at finding several (possibly overlapping) dense subgraphs which might correspond to communities in social networks or interesting events. In this talk we present a natural generalization of the densest subgraph problem, where the main goal is to find at most k subgraphs with maximum total aggregate density, while satisfying an upper bound on the pairwise Jaccard coefficient between the sets of nodes of the subgraphs. We will also illustrate how finding dense subgraphs can be an important subroutine for event detection in social media. Social media has great potential, as apart from the traditional media sources, many users post updates on different events. The highly dynamic nature of social networks gives the benefit of timely updates and the huge amount of content the benefit of diversity and large coverage. However, finding events presents also non-trivial challenges given the large amount of noisy and irrelevant data present in social media.