\nMy Phd student Aichetou Bouchareb (jointly advised by Marc Boull\u00e9<\/a> from Orange\nLabs) worked during her thesis on co-clustering mixed data. With Marc, we have\nbeen working since a few years on improvement and extension of his MODL\nframework. This is a fully automated non parametric density estimation\ntechnique based on grids which we extended in particular to graph clustering\nduring the Phd thesis of Romain Guigoures<\/a>. \n<\/p>\n\n \nIn Aichetou's work, the main goal was to enable true mixed data\nco-clustering. It is indeed quite easy to extend model based co-clustering to\nmixed data with both binary variables and numerical ones. However, variable\nclusters are not mixed in this case as they are either numerical or\nbinary. Aichetou defined an extension of MODL to the case where mixed clusters\nare possible. The key idea is to introduce a intermediate level of clustering:\nrather than clustering variables, we cluster so-called \"variable part\". A\nvariable part is a cluster of values<\/i> taken by a variable, for instance an\ninterval of values for a numerical variable. Once the value space of each\nvariable has been partitioned into variable parts, we can cluster those parts\nwith no type constraint. Aichetou defined a generative model based on this\nprinciple. She derived a prior distribution on its parameters and an\nestimation strategy based on the MAP principle. The strategy optimizes\neverything, including value spaces partitioning. \n<\/p>\n\n \nThis work is covered by the following publications:\n<\/p>\n \ntook place on the 28th of November. Aichetou gave an inspiring speech in front of\nthe following jury:\n<\/p>\n \nand myself.\n<\/p>\n\n \nThe summary of the thesis follows:\n<\/p>\n \nCo-clustering is a class of unsupervised data analysis techniques aiming at\nextracting the underlying dependency structure between the rows and columns of\na data table in the form of homogeneous blocks, known as co-clusters. These\ntechniques can be distinguished into those that aim at simultaneously\nclustering the instances and variables, and those that aim at clustering the\nvalues of two or more variables of a data set. Most of these techniques are\nlimited to variables of the same type, and are hardly scalable to large data\nsets while providing easily interpretable clusters and co-clusters. Among the\nexisting value based co-clustering approaches, MODL is suitable for processing\nlarge data sets with several numerical or categorical variables. In this\nthesis, we propose a value based approach, inspired by MODL, to perform a\nsimultaneous clustering of the instances and variables of a data set with\npotentially mixed-type variables. The proposed co-clustering model provides a\nMaximum A Posteriori based summary of the data that can be used as it is for\nexploratory analysis of the data. When the summary is large, exploratory\nanalysis tools, such as model coarsening, can be used to simplify the\nco-clustering which facilitates the interpretation of the results. We show\nthat the proposed co-clustering approach can handle large data and extract\neasily interpretable clusters from mixed data with more than 10 millions\nobservations. We also show the robustness of the approach, its capacity to\nextract inter-dependence between the variables, and its good behavior in\nextreme cases such as in the case of pattern-less data and in the case of\nperfectly correlated variables.\n<\/p>\nAichetou Bouchareb, \nA regularized approach of instances x variables co-clustering for exploratory data analysis<\/cite><\/small>\n<\/blockquote>\n\n \nThe thesis is available on TEL here<\/a>. \n<\/p>\n","summary":"In which Aichetou is the hero"},{"id":"https:\/\/apiacoa.org\/\/blog\/2018\/03\/welcome-to-elie-mengin.en","link":{"@attributes":{"type":"text\/html","rel":"alternate","href":"https:\/\/apiacoa.org\/\/blog\/2018\/03\/welcome-to-elie-mengin.en.html"}},"title":"Welcome to Elie Mengin","published":"2018-03-01T00:00:00+01:00","updated":"2018-03-01T00:00:00+01:00","content":" \nWe are welcoming in my research group\n(statistical learning and networks<\/a>) \na new Ph. D. student, Elie Mengin. He will be pursuing a Ph. D. in\ncollaboration Quarkslab<\/a> and under my supervision. \n<\/p>\n\n \nElie will be working on machine learning for computer security, with a\nparticular focus on two topics: static analysis of binaries and dynamic\nanalysis of instrumented programs. \n<\/p>\n\n \nYou can find more information about him on his\nlinkedin profile<\/a>. \n<\/p>\n","summary":"Where I welcome our new Ph.D. student, Elie Mengin"},{"id":"https:\/\/apiacoa.org\/\/blog\/2017\/11\/marco-corneli-phd-defense.en","link":{"@attributes":{"type":"text\/html","rel":"alternate","href":"https:\/\/apiacoa.org\/\/blog\/2017\/11\/marco-corneli-phd-defense.en.html"}},"title":"Marco Corneli's PhD Defense","published":"2017-11-17T00:00:00+01:00","updated":"2017-11-17T00:00:00+01:00","content":" \nMy Phd student Marco Corneli (jointly advised by\nPierre Latouche<\/a>)\nworked during his thesis on dynamic graphs. \n<\/p>\n\n \nThe core idea of the thesis was to avoid the snapshotting trick which is\ngenerally used to turn interaction data with continuous time stamps into a\ntime series of interaction graphs. While this is a nice way to switch to a\ndiscrete time view, one needs to have an idea of the natural time scale of the\ninteraction process. We need also to assume that this time scale is somewhat\nstationary. My first incursion in this non snapshotting idea was during the\nPhd thesis of \nRomain Guigoures<\/a>. \nWe studied at that time a variant of the MODL approach. In the present thesis,\nwe focused on generative models. \n<\/p>\n\n \nThe key idea was to consider that the flow of interactions between two actors\nfollows a non homogeneous Poisson point process. Then we assumed the intensity\nfunction to depend on pairwise hidden latent variables. In other words the\ntemporal behavior of the intensity between two actors depend only on the\nclasses of those actors.\n<\/p>\n\n \nUsing this idea as a guide, Marco was able to develop several models for\ndynamic graphs. Some of them still use a snapshotting trick, especially the\nquite advanced one that mixes text and dynamic graphs, but one of them achieve\nthe initial goal of the thesis by enabling us to analyze dynamic interactions\nin continuous time. \n<\/p>\n\n \nThis work is covered by the following publications:\n<\/p>\n \ntook place on the 17th of November. Marco gave a brilliant speech in front of\nthe following jury:\n<\/p>\n \nand myself.\n<\/p>\n\n \nThe summary of the thesis follows:\n<\/p>\n \nThe thesis is available on TEL here<\/a>. \n<\/p>\n","summary":"In which Marco is the hero"},{"id":"https:\/\/apiacoa.org\/\/blog\/2017\/10\/welcome-to-nicolas-jouvin.en","link":{"@attributes":{"type":"text\/html","rel":"alternate","href":"https:\/\/apiacoa.org\/\/blog\/2017\/10\/welcome-to-nicolas-jouvin.en.html"}},"title":"Welcome to Nicolas Jouvin","published":"2017-10-17T00:00:00+02:00","updated":"2017-10-17T00:00:00+02:00","content":" \nWe are welcoming in my research group (statistical learning and networks<\/a>) a\nnew Ph. D. student, Nicolas Jouvin. His grant is funded by the DIM-MathInnov<\/a>\nprogram from R\u00e9gion \u00cele-de-France, and he will work in collaboration with\nInstitut Curie, a specialized hospital in cancerology. Nicolas will be advised\nby Pierre Latouche<\/a>.\n<\/p>\n\n \nNicolas is from Paris and did most of his undergraduate studies in Paris 1\nPanth\u00e9on-Sorbonne university. He finished his postgraduate studies by a Master's\ndegree at the ENS Paris Saclay, majoring in Statistical Learning and a minor\nin Computer Vision. \n<\/p>\n\n \nFor his Ph. D., he will work on medical data in collaboration with\nanatomopatologist at the Institut Curie. The goal is to develop new techniques\nin order to segment an image, e.g. a biopsy, with the help of the text\nassociated to it, e.g. an anatomopatological report.\n<\/p>\n\n \nYou can find more information about him on his web page<\/a>.\n<\/p>\n","summary":"Where I welcome our new Ph.D. student, Antoine Lucquiaud"},{"id":"https:\/\/apiacoa.org\/\/blog\/2017\/10\/welcome-to-antoine-lucquiaud.en","link":{"@attributes":{"type":"text\/html","rel":"alternate","href":"https:\/\/apiacoa.org\/\/blog\/2017\/10\/welcome-to-antoine-lucquiaud.en.html"}},"title":"Welcome to Antoine Lucquiaud","published":"2017-10-17T00:00:00+02:00","updated":"2017-10-17T00:00:00+02:00","content":" \nWe are welcoming in my research group\n(statistical learning and networks<\/a>) a new Ph. D. student, Antoine Lucquiaud,\non an ENS Paris-Saclay scholarship. Antoine will be co-advised by \nJulien Randon-Furling<\/a> and myself. \n<\/p>\n\n \nAntoine was born in Lorient (Brittany), where he also attended high school and\nClasses Pr\u00e9paratoires aux Grandes \u00c9coles. He was then admitted to \u00c9cole\nNormale Sup\u00e9rieure Paris-Saclay (formerly ENS Cachan). He completed a Master's\ndegree in Complex Systems at Universit\u00e9 Pierre et Marie Curie, as well as a MA\nin Teaching for Physics and Chemistry (Agr\u00e9gation<\/i>). \n<\/p>\n\n \nHis PhD work investigates complex socio-economic systems, exploring in\nparticular the impact of random interventions on the global efficiency of such\nsystems. \n<\/p>\n","summary":"Where I welcome our new Ph.D. student, Antoine Lucquiaud"},{"id":"https:\/\/apiacoa.org\/\/blog\/2017\/09\/welcome-to-florian-barbaro.en","link":{"@attributes":{"type":"text\/html","rel":"alternate","href":"https:\/\/apiacoa.org\/\/blog\/2017\/09\/welcome-to-florian-barbaro.en.html"}},"title":"Welcome to Florian Barbaro","published":"2017-09-20T00:00:00+02:00","updated":"2017-09-20T00:00:00+02:00","content":" \nWe are welcoming in my research group (statistical learning and networks<\/a>) a\nnew Ph. D. student, Florian Barbaro. He will be pursuing a CIFRE Ph. D. under\nmy supervision in collaboration with BGFi Consulting, a subsidiary of ADNEOM<\/a>.\n<\/p>\n\n \nFlorian is originally from Corsica but received his education in Nice, notably\nat the Ecole Pr\u00e9paratoire aux Grandes \u00c9coles. He then completed his\nundergraduate studies at ECE Paris, majoring in financial engineering with a\nminor Big Data. \n<\/p>\n\n \nFor his Ph. D., he will try to quantify the impact of press releases on\nfinancial market using semantic analysis and machine learning. \n<\/p>\n\n \nYou can find more information about him on his\nlinkedin profile<\/a>. \n<\/p>\n","summary":"Where I welcome our new Ph.D. student, Forian Barbaro"},{"id":"https:\/\/apiacoa.org\/\/blog\/2017\/07\/humanit\u00e9s-et-statistiques.fr","link":{"@attributes":{"type":"text\/html","rel":"alternate","href":"https:\/\/apiacoa.org\/\/blog\/2017\/07\/humanit%C3%A9s-et-statistiques.fr.html"}},"title":"Humanit\u00e9s et Statistiques","published":"2017-07-07T00:00:00+02:00","updated":"2017-07-07T00:00:00+02:00","content":" \nApr\u00e8s un long travail \u00e9ditorial, j'ai le plaisir d'annoncer la publication\ndu Journal de la Soci\u00e9t\u00e9 Fran\u00e7aise de Statistiques<\/a> sur\nle th\u00e8me \u00ab Humanit\u00e9s et statistiques \u00bb<\/a>.\n<\/p>\n\n \n \nEn vous souhaitant une bonne lecture !\n<\/p>\n","summary":"Num\u00e9ro sp\u00e9cial \u00ab Humanit\u00e9s et Statistiques \u00bb"},{"id":"https:\/\/apiacoa.org\/\/blog\/2017\/07\/humanities-and-statistics.en","link":{"@attributes":{"type":"text\/html","rel":"alternate","href":"https:\/\/apiacoa.org\/\/blog\/2017\/07\/humanities-and-statistics.en.html"}},"title":"Humanities and Statistics","published":"2017-07-07T00:00:00+02:00","updated":"2017-07-07T00:00:00+02:00","content":" \nAfter a long editorial work, I'm pleased to announced the publication of a\nspecial issue of the\nJournal de la Soci\u00e9t\u00e9 Fran\u00e7aise de Statistiques<\/a> \nabout \"Humanities and statistics\"<\/a>.\n<\/p>\n\n \n \nWe hope you will enjoy reading the special issue as much as we enjoy editing\nit. \n<\/p>\n","summary":"Special issue about \"Humanities and statistics\""},{"id":"https:\/\/apiacoa.org\/\/blog\/2017\/02\/conan-les-relances.fr","link":{"@attributes":{"type":"text\/html","rel":"alternate","href":"https:\/\/apiacoa.org\/\/blog\/2017\/02\/conan-les-relances.fr.html"}},"title":"Conan : les relances","published":"2017-02-04T00:00:00+01:00","updated":"2017-02-04T00:00:00+01:00","content":" \nApr\u00e8s une premi\u00e8re analyse simple<\/a> des combats dans le jeu Conan<\/a>,\nj'approfondis ici l'analyse en tenant compte des possibilit\u00e9s de relance des\nd\u00e9s. Mais avant, revenons au combat d'Hadrathus et Zogar Sag…\n<\/p>\n\n\n \nDans la note pr\u00e9c\u00e9dente, nous avions laiss\u00e9 notre sorcier Hadrathus bien\nd\u00e9sempar\u00e9 en face de Zogar Sag, avec une probabilit\u00e9 estim\u00e9e rapidement \u00e0 1\nsur 1 million de tuer son ennemi (\u00e0 la dague, une id\u00e9e qui semble assez peu\nstrat\u00e9gique, quand on y r\u00e9fl\u00e9chit bien…). \n<\/p>\n\n\n \nRevenons sur les strat\u00e9gies simples o\u00f9 Hadrathus reste actif \u00e0 tous les\ntours et d\u00e9pense les deux gemmes qu'il r\u00e9cup\u00e8re en d\u00e9but de chaque\ntour. Ceci lui permet de faire deux attaques \u00e0 deux d\u00e9s ou une attaque \u00e0\ntrois d\u00e9s \u00e0 chaque tour. En tenant compte de la d\u00e9fense passive de 3 de\nZogar Sag, les probabilit\u00e9s de causer diff\u00e9rents niveaux de d\u00e9g\u00e2ts sont\ndonn\u00e9es dans la table suivante :\n<\/p>\n\n
The defense<\/h2>\n
\n
\n
Generative models for Dynamic Graphs<\/h2>\n
\n
The defense<\/h2>\n
\n
\n Graphs are mathematical structures very suitable to model interactions\n between objects or actors of interest. Several real networks such as\n communication networks, financial transaction networks, mobile telephone\n networks and social networks (Facebook, Linkedin, etc.) can be modeled via\n graphs. When observing a network, the time variable comes into play in two\n different ways: we can study the time dates at which the interactions occur\n and\/or the interaction time spans. This thesis only focuses on the first time\n dimension and each interaction is assumed to be instantaneous, for\n simplicity. Hence, the network evolution is given by the interaction time\n dates only. In this framework, graphs can be used in two different ways to\n model networks: 1) Discrete time. A network is observed at several times and\n a graph is associated with each observation time. Two nodes of a graph are\n connected if one or more interactions occurred between them in the\n corresponding time frame. Thus, interactions are aggregated between two\n consecutive observation times and the exact interaction dates are lost. In\n this context, a dynamic network is represented by a sequence of graphs. 2)\n Continuous Time. Several edges are allowed to connect the nodes of a graph at\n different times. One edge is uniquely associated with a pair of nodes and a\n time point. No aggregation is required and interaction times are never\n lost. Therefore, a dynamic network is represented by a single multiple graph\n whose edges are labeled by the interaction times. In this thesis both these\n perspectives are adopted, alternatively. We consider new unsupervised methods\n to cluster the nodes of a graph into groups of homogeneous connection\n profiles. In this manuscript, the node groups are assumed to be time\n invariant to avoid possible identifiability issues. Moreover, the approaches\n that we propose aim to detect structural changes in the way the node clusters\n interact with each other. The building block of this thesis is the stochastic\n block model (SBM), a probabilistic approach initially used in social\n sciences. The standard SBM assumes that the nodes of a graph belong to hidden\n (disjoint) clusters and that the probability of observing an edge between two\n nodes only depends on their clusters. Since no further assumption is made on\n the connection probabilities, SBM is a very flexible model able to detect\n different network topologies (hubs, stars, communities, etc.). By adapting\n the block modeling perspective of SBM to dynamic graphs, the main\n contributions of this thesis are the following: 1. We introduce a new\n extension of SBM for dynamic graphs. The proposed approach, called dSBM,\n adopts non homogeneous Poisson processes to model the interaction times\n between pairs of nodes in dynamic graphs, either in discrete or continuous\n time. The intensity functions of the processes only depend on the node\n clusters, in a block modeling perspective. Moreover, all the intensity\n functions share some regularity properties on hidden time intervals that need\n to be estimated. 2. A recent estimation algorithm for SBM, based on the\n greedy maximization of an exact criterion (exact ICL) is adopted for\n inference and model selection in dSBM. To the best of our knowledge, this is\n the first time this algorithm is adopted for inference in dynamic stochastic\n block models. 3. An exact algorithm for change point detection in time\n series, the \"pruned exact linear time\" (PELT) method is extended to deal with\n dynamic graph data modeled via dSBM. The approach we propose can be used for\n change point analysis in graph data. 4. A further extension of dSBM is\n developed to analyze dynamic networks with textual edges (like social\n networks, for instance). In this context, the graph edges are associated with\n documents exchanged between the corresponding nodes. The textual content of\n the documents can provide additional information about the dynamic graph\n topological structure. The new model we propose is called \"dynamic stochastic\n topic block model\" (dSTBM). This manuscript is organized as follows. In the\n first chapter, we pass through the main notions of graph theory and review\n some stylized facts about real networks. Two formal definitions of dynamic\n graph are provided. Then, the main existing generative models for static and\n dynamic random graphs are presented along with their associated inference\n procedures. Finally, some statistical tools not necessarily related with\n network analysis are described in detail since they are used in later\n chapters. In the second chapter, two versions of dSBM are introduced, both\n dealing with discrete time dynamic graphs. The corresponding inference\n procedure aims to maximize the complete data integrated log-likelihood, thus\n allowing us to learn the model parameters and select the number of clusters\n at the same time. In the third chapter, we model continuous time dynamic\n graphs via dSBM and focus on clustering and change point analysis in graph\n data. A standard variational approach is adopted for the inference and one\n step of the estimation algorithm relies on the PELT method. Finally, the\n fourth chapter introduces the dSTBM for discrete time dynamic graph with\n textual edges. The inference procedure is detailed and a model selection\n criterion is formally obtained. The last part of each chapter is devoted to\n experiments on both simulated and real data. These experiments allow us to\n highlight the features of the proposed approaches and to compare them with\n alternative methods. \nMarco Corneli, Dynamic stochastic block models,\nclustering and segmentation in dynamic graphs\n<\/cite><\/small>\n<\/blockquote>\n\n
Num\u00e9ro sp\u00e9cial \u00ab Humanit\u00e9s et Statistiques \u00bb<\/h2>\n
\"Humanities and statistics\" special issue<\/h2>\n
Hadrathus contre Zogar Sag, le retour<\/h2>\n
Strat\u00e9gies simples<\/h3>\n