Defended PhD Theses

Cette dissertation apporte un éclairage nouveau sur les capacités des méthodes d'apprentissage automatique. Nous démontrons d'abord que les attaques profilées paramétriques surpassent les méthodes d'apprentissage automatique lorsqu'il n'y a pas d'erreur d'estimation ni d'hypoth`ese. En revanche, les attaques fondées sur l'apprentissage automatique sont avantageuses dans des scénarios réalistes o`u le nombre de données lors de l'étape d'apprentissage est faible. Par la suite, nous proposons une nouvelle métrique formelle d'évaluation qui permet (1) de comparer des attaques paramétriques et non-paramétriques et (2) d'interpréter les résultats de chaque méthode. La nouvelle mesure fournit les causes d'un taux de réussite élevé ou faible d'une attaque et, par conséquent, donne des pistes pour améliorer l'évaluation d'une implantation. Enfin, nous présentons des résultats expérimentaux sur des appareils non protégés et protégés. La premi`ere étude montre que l'apprentissage automatique a un taux de réussite plus élevé qu'une méthode paramétrique lorsque seules quelques données sont disponibles. La deuxi`eme expérience démontre qu'un dispositif protégé est attaquable avec une approche appartenant `a l'apprentissage automatique. La stratégie basée sur l'apprentissage automatique nécessite le m^eme nombre de données lors de la phase d'apprentissage que lorsque celle-ci attaque un produit non protégé. Nous montrons également que des méthodes paramétriques surestiment ou sous-estiment le niveau de sécurité fourni par l'appareil alors que l'approche basée sur l'apprentissage automatique améliore cette estimation.

En résumé, notre th`ese est que les attaques basées sur l'apprentissage automatique sont avantageuses par rapport aux techniques classiques lorsque la quantité d'information a priori sur l'appareil cible et le nombre de données lors de la phase d'apprentissage sont faibles.},
note = {Funder: Universite Libre de Bruxelles},
keywords = {},
pubstate = {published},
tppubtype = {phdthesis}
}

Historically, time series forecasting has been mainly studied in econometrics and statistics. In the last two decades, machine learning, a field that is concerned with the development of algorithms that can automatically learn from data, has become one of the most active areas of predictive modeling research. This success is largely due to the superior performance of machine learning prediction algorithms in many different applications as diverse as natural language processing, speech recognition and spam detection. However, there has been very little research at the intersection of time series forecasting and machine learning.

The goal of this dissertation is to narrow this gap by addressing the problem of multi-step-ahead time series forecasting from the perspective of machine learning. To that end, we propose a series of forecasting strategies based on machine learning algorithms.

Multi-step-ahead forecasts can be produced recursively by iterating a one-step-ahead model, or directly using a specific model for each horizon. As a first contribution, we conduct an in-depth study to compare recursive and direct forecasts generated with different learning algorithms for different data generating processes. More precisely, we decompose the multi-step mean squared forecast errors into the bias and variance components, and analyze their behavior over the forecast horizon for different time series lengths. The results and observations made in this study then guide us for the development of new forecasting strategies.

In particular, we find that choosing between recursive and direct forecasts is not an easy task since it involves a trade-off between bias and estimation variance that depends on many interacting factors, including the learning model, the underlying data generating process, the time series length and the forecast horizon. As a second contribution, we develop multi-stage forecasting strategies that do not treat the recursive and direct strategies as competitors, but seek to combine their best properties. More precisely, the multi-stage strategies generate recursive linear forecasts, and then adjust these forecasts by modeling the multi-step forecast residuals with direct nonlinear models at each horizon, called rectification models. We propose a first multi-stage strategy, that we called the rectify strategy, which estimates the rectification models using the nearest neighbors model. However, because recursive linear forecasts often need small adjustments with real-world time series, we also consider a second multi-stage strategy, called the boost strategy, that estimates the rectification models using gradient boosting algorithms that use so-called weak learners.

Generating multi-step forecasts using a different model at each horizon provides a large modeling flexibility. However, selecting these models independently can lead to irregularities in the forecasts that can contribute to increase the forecast variance. The problem is exacerbated with nonlinear machine learning models estimated from short time series. To address this issue, and as a third contribution, we introduce and analyze multi-horizon forecasting strategies that exploit the information contained in other horizons when learning the model for each horizon. In particular, to select the lag order and the hyperparameters of each model, multi-horizon strategies minimize forecast errors over multiple horizons rather than just the horizon of interest.

We compare all the proposed strategies with both the recursive and direct strategies. We first apply a bias and variance study, then we evaluate the different strategies using real-world time series from two past forecasting competitions. For the rectify strategy, in addition to avoiding the choice between recursive and direct forecasts, the results demonstrate that it has better, or at least has close performance to, the best of the recursive and direct forecasts in different settings. For the multi-horizon strategies, the results emphasize the decrease in variance compared to single-horizon strategies, especially with linear or weakly nonlinear data generating processes. Overall, we found that the accuracy of multi-step-ahead forecasts based on machine learning algorithms can be significantly improved if an appropriate forecasting strategy is used to select the model parameters and to generate the forecasts.

Lastly, as a fourth contribution, we have participated in the Load Forecasting track of the Global Energy Forecasting Competition 2012. The competition involved a hierarchical load forecasting problem where we were required to backcast and forecast hourly loads for a US utility with twenty geographical zones. Our team, TinTin, ranked fifth out of 105 participating teams, and we have been awarded an IEEE Power & Energy Society award.

},
note = {Funder: Universite Libre de Bruxelles},
keywords = {},
pubstate = {published},
tppubtype = {phdthesis}
}

Another important problem in bioinformatics is the question of how the inferred networks' quality can be evaluated. The current best practice is a two step procedure. In the first step, the highest scoring interactions are compared to known interactions stored in biological databases. The inferred networks passes this quality assessment if there is a large overlap with the known interactions. In this case, a second step is carried out in which unknown but high scoring and thus promising new interactions are validated 'by hand' via laboratory experiments. Unfortunately when integrating prior knowledge in the inference procedure, this validation procedure would be biased by using the same information in both the inference and the validation. Therefore, it would no longer allow an independent validation of the resulting network.

The main contribution of this thesis is a complete computational framework that uses experimental knock down data in a cross-validation scheme to both infer and validate directed networks. Its components are i) a method that integrates genomic data and prior knowledge to infer directed networks, ii) its implementation in an R/Bioconductor package and iii) a web application to retrieve prior knowledge from PubMed abstracts and biological databases. To infer directed networks from genomic data and prior knowledge, we propose a two step procedure: First, we adapt the pairwise feature selection strategy mRMR to integrate prior knowledge in order to obtain the network's skeleton. Then for the subsequent orientation phase of the algorithm, we extend a criterion based on interaction information to include prior knowledge. The implementation of this method is available both as part of the prior retrieval tool Predictive Networks and as a stand-alone R/Bioconductor package named predictionet.

Furthermore, we propose a fully data-driven quantitative validation of such directed networks using experimental knock-down data: We start by identifying the set of genes that was truly affected by the perturbation experiment. The rationale of our validation procedure is that these truly affected genes should also be part of the perturbed gene's childhood in the inferred network. Consequently, we can compute a performance score},
note = {Funder: Universite Libre de Bruxelles},
keywords = {},
pubstate = {published},
tppubtype = {phdthesis}
}

massive amounts of streaming data.

In many applications, data is collected for modeling the processes. The process model is hoped to drive objectives such as decision support, data visualization, business intelligence, automation and control, pattern recognition and classification, etc. However, we face significant challenges in data-driven modeling of processes. Apart from the errors, outliers and noise in the data measurements, the main challenge is due to a large dimensionality, which is the number of variables each data sample measures. The samples often form a long temporal sequence called a multivariate time series where any one sample is influenced by the others.

We wish to build a model that will ensure robust generation, reviewing, and representation of new multivariate time series that are consistent with the underlying process.

In this thesis, we adopt a modeling framework to extract characteristics from multivariate time series that correspond to dynamic variation-covariation common to the measured variables across all the samples. Those characteristics of a multivariate time series are named its 'commonalities' and a suitable measure for them is defined. What makes the multivariate time series model versatile is the assumption regarding the existence of a latent time series of known or presumed characteristics and much lower dimensionality than the measured time series; the result is the well-known 'dynamic factor model'.

Original variants of existing methods for estimating the dynamic factor model are developed: The estimation is performed using the frequency-domain equivalent of the dynamic factor model named the 'spectral factor model'. To estimate the spectral factor model, ideas are sought from the asymptotic theory of spectral estimates. This theory is used to attain a probabilistic formulation, which provides maximum likelihood estimates for the spectral factor model parameters. Then, maximum likelihood parameters are developed with all the analysis entirely in the spectral-domain such that the dynamically transformed latent time series inherits the commonalities maximally.

The main contribution of this thesis is a learning framework using the spectral factor model. We term learning as the ability of a computational model of a process to robustly characterize the data the process generates for purposes of pattern matching, classification and prediction. Hence, the spectral factor model could be claimed to have learned a multivariate time series if the latent time series when dynamically transformed extracts the commonalities reliably and maximally. The spectral factor model will be used for mainly two multivariate time series learning applications: First, real-world streaming datasets obtained from various processes are to be classified; in this exercise, human brain magnetoencephalography signals obtained during various cognitive and physical tasks are classified. Second, the commonalities are put to test by asking for reliable prediction of a multivariate time series given its past evolution; share prices in a portfolio are forecasted as part of this challenge.

For both spectral factor modeling and learning, an analytical solution as well as an iterative solution are developed. While the analytical solution is based on low-rank approximation of the spectral density function, the iterative solution is based on the expectation-maximization algorithm. For the human brain signal classification exercise, a strategy for comparing similarities between the commonalities for various classes of multivariate time series processes is developed. For the share price prediction problem, a vector autoregressive model whose parameters are enriched with the maximum likelihood commonalities is designed. In both these learning problems, the spectral factor model gives commendable performance with respect to competing approaches.

Les processus informatisés actuels gén`erent des quantités massives de flux de données. Dans nombre d'applications, ces flux de données sont collectées en vue de modéliser les processus. Les mod`eles de processus obtenus ont pour but la réalisation d'objectifs tels que l'aide `a la décision, la visualisation de données, l'informatique décisionnelle, l'automatisation et le contr^ole, la reconnaissance de formes et la classification, etc. La modélisation de processus sur la base de données implique cependant de faire face `a d'importants défis. Outre les erreurs, les données aberrantes et le bruit, le principal défi provient de la large dimensionnalité, i.e. du nombre de variables dans chaque échantillon de données mesurées. Les échantillons forment souvent une longue séquence temporelle appelée série temporelle multivariée, o`u chaque échantillon est influencé par les autres. Notre objectif est de construire un mod`ele robuste qui garantisse la génération, la révision et la représentation de nouvelles séries temporelles multivariées cohérentes avec le processus sous-jacent.

Dans cette th`ese, nous adoptons un cadre de modélisation capable d'extraire, `a partir de séries temporelles multivariées, des caractéristiques correspondant `a des variations - covariations dynamiques communes aux variables mesurées dans tous les échantillons. Ces caractéristiques sont appelées «points communs» et une mesure qui leur est appropriée est définie. Ce qui rend le mod`ele de séries temporelles multivariées polyvalent est l'hypoth`ese relative `a l'existence de séries temporelles latentes de caractéristiques connues ou présumées et de dimensionnalité beaucoup plus faible que les séries temporelles mesurées; le résultat est le bien connu «mod`ele factoriel dynamique». Des variantes originales de méthodes existantes pour estimer le mod`ele factoriel dynamique sont développées : l'estimation est réalisée en utilisant l'équivalent du mod`ele factoriel dynamique au niveau du domaine de fréquence, désigné comme le «mod`ele factoriel spectral». Pour estimer le mod`ele factoriel spectral, nous nous basons sur des idées relatives `a la théorie des estimations spectrales. Cette théorie est utilisée pour aboutir `a une formulation probabiliste, qui fournit des estimations de probabilité maximale pour les param`etres du mod`ele factoriel spectral. Des param`etres de probabilité maximale sont alors développés, en plac cant notre analyse enti`erement dans le domaine spectral, de fac con `a ce que les séries temporelles latentes transformées dynamiquement héritent au maximum des points communs.

La principale contribution de cette th`ese consiste en un cadre d'apprentissage utilisant le mod`ele factoriel spectral. Nous désignons par apprentissage la capacité d'un mod`ele de processus `a caractériser de fac con robuste les données générées par le processus `a des fins de filtrage par motif, classification et prédiction. Dans ce contexte, le mod`ele factoriel spectral est considéré comme ayant appris une série temporelle multivariée si la série temporelle latente, une fois dynamiquement transformée, permet d'extraire les points communs de fac con fiable et maximale. Le mod`ele factoriel spectral sera utilisé principalement pour deux applications d'apprentissage de séries multivariées : en premier lieu, des ensembles de données sous forme de flux venant de différents processus du monde réel doivent ^etre classifiés; lors de cet exercice, la classification porte sur des signaux magnétoencéphalographiques obtenus chez l'homme au cours de différentes t^aches physiques et cognitives; en second lieu, les points communs obtenus sont testés en demandant une prédiction fiable d'une série temporelle multivariée étant donnée l'évolution passée; les prix d'un portefeuille d'actions sont prédits dans le cadre de ce défi.

`A la fois pour la modélisation et pour l'apprentissage factoriel spectral, une solution analytique aussi bien qu'une solution itérative sont développées. Tandis que la solution analytique est basée sur une approximation de rang inférieur de la fonction de densité spectrale, la solution itérative est basée, quant `a elle, sur l'algorithme de maximisation des attentes. Pour l'exercice de classification des signaux magnétoencéphalographiques humains, une stratégie de comparaison des similitudes entre les points communs des différentes classes de processus de séries temporelles multivariées est développée. Pour le probl`eme de prédiction des prix des actions, un mod`ele vectoriel autorégressif dont les param`etres sont enrichis avec les points communs de probabilité maximale est conc cu. Dans ces deux probl`emes d'apprentissage, le mod`ele factoriel spectral atteint des performances louables en regard d'approches concurrentes.},
note = {Funder: Universite Libre de Bruxelles},
keywords = {},
pubstate = {published},
tppubtype = {phdthesis}
}

Dans un premier cas, les tests viseront `a maximiser la somme des gains collectés. Un juste compromis doit alors ^etre trouvé entre l'exploitation et l'exploration. Ce probl`eme est couramment dénommé dans la littérature scientifique "multi-armed bandit problem".

Dans un second cas, un nombre de sélections maximal est imposé et l'objectif consistera `a répartir ces sélections de fac con `a augmenter les chances de trouver l'alternative présentant le gain moyen le plus élevé. Ce deuxi`eme probl`eme est couramment repris dans la littérature scientifique sous l'appellation "selecting the best".

La sélection de type gloutonne joue un r^ole important dans la résolution de ces probl`emes de décision et op`ere en choisissant l'alternative qui s'est jusqu'ici montrée optimale. Or, la nature généralement aléatoire de l'environnement rend incertains les résultats d'une telle sélection.

Dans cette th`ese, nous introduisons une nouvelle quantité, appelée le "gain espéré d'une action gloutonne". Sur base de quelques propriétés de cette quantité, de nouveaux algorithmes permettant de résoudre les deux probl`emes décisionnels précités seront proposés.

Une attention particuli`ere sera ici pr^etée `a l'application des techniques présentées au domaine de la sélection de mod`eles en l'apprentissage artificiel supervisé.

La collaboration avec le service d'anesthésie de l'H^opital Erasme nous a permis d'appliquer les algorithmes proposés `a des données réelles, provenant du milieu médical. Nous avons également développé un syst`eme d'aide `a la décision dont un prototype a déj`a été testé en conditions réelles sur un échantillon restreint de patients.},
note = {Funder: Universite Libre de Bruxelles},
keywords = {},
pubstate = {published},
tppubtype = {phdthesis}
}

Unfortunately, even for small model organisms such as the yeast Saccharomyces cerevisiae, the number p of genes is much larger than the number n of expression data samples. The dimensionality issue induced by this ``small n, large p' data setting renders standard statistical learning methods inadequate. Restricting the complexity of the models enables to deal with this serious impediment. Indeed, by introducing (a priori undesirable) bias in the model selection procedure, one reduces the variance of the selected model thereby increasing its accuracy.

Gaussian graphical models (GGMs) have proven to be a very powerful formalism to infer GRNs from expression data. Standard GGM selection techniques can unfortunately not be used in the ``small n, large p' data setting. One way to overcome this issue is to resort to regularization. In particular, shrinkage estimators of the covariance matrix--required to infer GGMs--have proven to be very effective. Our first contribution consists in a new shrinkage estimator that improves upon existing ones through the use of a Monte Carlo (parametric bootstrap) procedure.

Another approach to GGM selection in the ``small n, large p' data setting consists in reverse engineering limited-order partial correlation graphs (q-partial correlation graphs) to approximate GGMs. Our second contribution consists in an inference algorithm, the q-nested procedure, that builds a sequence of nested q-partial correlation graphs to take advantage of the smaller order graphs' topology to infer higher order graphs. This allows us to significantly speed up the inference of such graphs and to avoid problems related to multiple testing. Consequently, we are able to consider higher order graphs, thereby increasing the accuracy of the inferred graphs.

Another important challenge in bioinformatics is the prediction of gene function. An example of such a prediction task is the identification of genes that are targets of the nitrogen catabolite repression (NCR) selection mechanism in the yeast Saccharomyces cerevisiae. The study of model organisms such as Saccharomyces cerevisiae is indispensable for the understanding of more complex organisms. Our third contribution consists in extending the standard two-class classification approach by enriching the set of variables and comparing several feature selection techniques and classification algorithms.

Finally, our fourth contribution formulates the prediction of NCR target genes as a network inference task. We use GGM selection to infer multivariate dependencies between genes, and, starting from a set of genes known to be sensitive to NCR, we classify the remaining genes. We hence avoid problems related to the choice of a negative training set and take advantage of the robustness of GGM selection techniques in the ``small n, large p' data setting.},
note = {Funder: Universite Libre de Bruxelles},
keywords = {},
pubstate = {published},
tppubtype = {phdthesis}
}

In environmental monitoring studies, many applications are expected to run unattended for months or years. Sensor nodes are however constrained by limited resources, particularly in terms of energy. Since communication is one order of magnitude more energy-consuming than processing, the design of data collection schemes that limit the amount of transmitted data is therefore recognized as a central issue for wireless sensor networks.

An efficient way to address this challenge is to approximate, by means of mathematical models, the evolution of the measurements taken by sensors over space and/or time. Indeed, whenever a mathematical model may be used in place of the true measurements, significant gains in communications may be obtained by only transmitting the parameters of the model instead of the set of real measurements. Since in most cases there is little or no a priori information about the variations taken by sensor measurements, the models must be identified in an automated manner. This calls for the use of machine learning techniques, which allow to model the variations of future measurements on the basis of past measurements.

This thesis brings two main contributions to the use of learning techniques in a sensor network. First, we propose an approach which combines time series prediction and model selection for reducing the amount of communication. The rationale of this approach, called adaptive model selection, is to let the sensors determine in an automated manner a prediction model that does not only fits their measurements, but that also reduces the amount of transmitted data.

The second main contribution is the design of a distributed approach for modeling sensed data, based on the principal component analysis (PCA). The proposed method allows to transform along a routing tree the measurements taken in such a way that (i) most of the variability in the measurements is retained, and (ii) the network load sustained by sensor nodes is reduced and more evenly distributed, which in turn extends the overall network lifetime. The framework can be seen as a truly distributed approach for the principal component analysis, and finds applications not only for approximated data collection tasks, but also for event detection or recognition tasks.

/

Les réseaux de capteurs sans fil forment une nouvelle famille de syst`emes informatiques permettant d'observer le monde avec une résolution sans précédent. En particulier, ces syst`emes promettent de révolutionner le domaine de l'étude environnementale. Un tel réseau est composé d'un ensemble de capteurs sans fil, ou unités sensorielles, capables de collecter, traiter, et transmettre de l'information. Gr^ace aux avancées dans les domaines de la microélectronique et des technologies sans fil, ces syst`emes sont `a la fois peu volumineux et peu co^uteux. Ceci permet leurs deploiements dans différents types d'environnements, afin d'observer l'évolution dans le temps et l'espace de quantités physiques telles que la température, l'humidité, la lumi`ere ou le son.

Dans le domaine de l'étude environnementale, les syst`emes de prise de mesures doivent souvent fonctionner de mani`ere autonome pendant plusieurs mois ou plusieurs années. Les capteurs sans fil ont cependant des ressources limitées, particuli`erement en terme d'énergie. Les communications radios étant d'un ordre de grandeur plus co^uteuses en énergie que l'utilisation du processeur, la conception de méthodes de collecte de données limitant la transmission de données est devenue l'un des principaux défis soulevés par cette technologie.

Ce défi peut ^etre abordé de mani`ere efficace par l'utilisation de mod`eles mathématiques modélisant l'évolution spatiotemporelle des mesures prises par les capteurs. En effet, si un tel mod`ele peut ^etre utilisé `a la place des mesures, d'importants gains en communications peuvent ^etre obtenus en utilisant les param`etres du mod`ele comme substitut des mesures. Cependant, dans la majorité des cas, peu ou aucune information sur la nature des mesures prises par les capteurs ne sont disponibles, et donc aucun mod`ele ne peut ^etre a priori défini. Dans ces cas, les techniques issues du domaine de l'apprentissage machine sont particuli`erement appropriées. Ces techniques ont pour but de créer ces mod`eles de fac con autonome, en anticipant les mesures `a venir sur la base des mesures passées.

Dans cette th`ese, deux contributions sont principalement apportées permettant l'applica-tion de techniques d'apprentissage machine dans le domaine des réseaux de capteurs sans fil. Premi`erement, nous proposons une approche qui combine la prédiction de série temporelle avec la sélection de mod`eles afin de réduire la communication. La logique de cette approche, appelée sélection de mod`ele adaptive, est de permettre aux unités sensorielles de determiner de mani`ere autonome un mod`ele de prédiction qui anticipe correctement leurs mesures, tout en réduisant l'utilisation de leur radio.

Deuxi`emement, nous avons conc cu une méthode permettant de modéliser de fac con distribuée les mesures collectées, qui se base sur l'analyse en composantes principales (ACP). La méthode permet de transformer les mesures le long d'un arbre de routage, de fac con `a ce que (i) la majeure partie des variations dans les mesures des capteurs soient conservées, et (ii) la charge réseau soit réduite et mieux distribuée, ce qui permet d'augmenter également la durée de vie du réseau. L'approche proposée permet de véritablement distribuer l'ACP, et peut ^etre utilisée pour des applications impliquant la collecte de données, mais également pour la détection ou la classification d'événements.

},
note = {Funder: Universite Libre de Bruxelles},
keywords = {},
pubstate = {published},
tppubtype = {phdthesis}
}

The traditional approach to evaluating breast cancer prognosis is based on the assessment of clinico-pathologic factors known to be associated with breast cancer survival. These factors are used to make recommendations about whether further treatment is required after the removal of a tumor by surgery. Treatment such as chemotherapy depends on the estimation of patients' risk of relapse. Although current approaches do provide good prognostic assessment of breast cancer survival, clinicians are aware that there is still room for improvement in the accuracy of their prognostic estimations.

In the late nineties, new high throughput technologies such as the gene expression profiling through microarray technology emerged. Microarrays allowed scientists to analyze for the first time the expression of the whole human genome ("transcriptome"). It was hoped that the analysis of genome-wide molecular data would bring new insights into the critical, underlying biological mechanisms involved in breast cancer progression, as well as significantly improve prognostic prediction. However, the analysis of microarray data is a difficult task due to their intrinsic characteristics: (i) thousands of gene expressions are measured for only few samples; (ii) the measurements are usually "noisy"; and (iii) they are highly correlated due to gene co-expressions. Since traditional statistical methods were not adapted to these settings, machine learning methods were picked up as good candidates to overcome these difficulties. However, applying machine learning methods for microarray analysis involves numerous steps, and the results are prone to overfitting. Several authors have highlighted the major pitfalls of this process in the early publications, shedding new light on the promising but overoptimistic results.

Since 2002, large comparative studies have been conducted in order to identify the key characteristics of successful methods for class discovery and classification. Yet methods able to identify robust molecular signatures that can predict breast cancer prognosis have been lacking. To fill this important gap, this thesis presents an original methodology dealing specifically with the analysis of microarray and survival data in order to build prognostic models and provide an honest estimation of their performance. The approach used for signature extraction consists of a set of original methods for feature transformation, feature selection and prediction model building. A novel statistical framework is presented for performance assessment and comparison of risk prediction models.

In terms of applications, we show that these methods, used in combination with a priori biological knowledge of breast cancer and numerous public microarray datasets, have resulted in some important discoveries. In particular, the research presented here develops (i) a robust model for the identification of breast molecular subtypes and (ii) a new prognostic model that takes into account the molecular heterogeneity of breast cancers observed previously, in order to improve traditional clinical guidelines and state-of-the-art gene signatures./Cette th`ese concerne le développement de techniques d'apprentissage (machine learning) afin de mettre au point de nouveaux outils cliniques basés sur des données moleculaires. Nous avons focalisé notre recherche sur le cancer du sein, un des cancers les plus fréquemment diagnostiqués. Ces outils sont développés dans le but d'aider les médecins dans leur évaluation du devenir clinique des patients cancéreux (cf. le pronostique).

Les approches traditionnelles d'évaluation du pronostique d'un patient cancéreux se base sur des crit`eres clinico-pathologiques connus pour ^etre prédictifs de la survie. Cette évaluation permet aux médecins de décider si un traitement est nécessaire apr`es l'extraction de la tumeur. Bien que les outils d'évaluation traditionnels sont d'une aide importante, les cliniciens sont conscients de la nécessité d'améliorer de tels outils.

Dans les années 90, de nouvelles technologies `a haut-débit, telles que le profilage de l'expression génique par biopuces `a ADN (microarrays), ont été mises au point afin de permettre aux scientifiques d'analyser l'expression de l'enti`ereté du gén^ome de cellules cancéreuses. Ce nouveau type de données moléculaires porte l'espoir d'améliorer les outils pronostiques traditionnels et d'approfondir nos connaissances concernant la gén`ese du cancer du sein. Cependant ces données sont extr^emement difficiles `a analyser `a cause (i) de leur haute dimensionalité (plusieurs dizaines de milliers de g`enes pour seulement quelques centaines d'expériences); (ii) du bruit important dans les mesures; (iii) de la collinéarité entre les mesures d^ue `a la co-expression des g`enes.

Depuis 2002, des études comparatives `a grande échelle ont permis d'identifier les méthodes performantes pour l'analyse de groupements et la classification de données microarray, négligeant l'analyse de survie pertinente pour le pronostique dans le cancer du sein. Pour pallier ce manque, cette th`ese présente une méthodologie originale adaptée `a l'analyse de données microarray et de survie afin de construire des mod`eles pronostiques performants et robustes.

En termes d'applications, nous montrons que cette méthodologie, utilisée en combinaison avec des connaissances biologiques a priori et de nombreux ensembles de données publiques, a permis d'importantes découvertes. En particulier, il résulte de la recherche presentée dans cette th`ese, le développement d'un mod`ele robuste d'identification des sous-types moléculaires du cancer du sein et de plusieurs signatures géniques améliorant significativement l'état de l'art au niveau pronostique.},
note = {Funder: Universite Libre de Bruxelles},
keywords = {},
pubstate = {published},
tppubtype = {phdthesis}
}

data. In a lot of emerging fields, like bioinformatics, they are confronted with datasets

having thousands of variables, a lot of noise, non-linear dependencies and, only, tens of

samples. The detection of functional relationships, when such uncertainty is contained in

data, constitutes a major challenge.

Our work focuses on variable selection and network inference from datasets having

many variables and few samples (high variable-to-sample ratio), such as microarray data.

Variable selection is the topic of machine learning whose objective is to select, among a

set of input variables, those that lead to the best predictive model. The application of

variable selection methods to gene expression data allows, for example, to improve cancer

diagnosis and prognosis by identifying a new molecular signature of the disease. Network

inference consists in representing the dependencies between the variables of a dataset by

a graph. Hence, when applied to microarray data, network inference can reverse-engineer

the transcriptional regulatory network of cell in view of discovering new drug targets to

cure diseases.

In this work, two original tools are proposed MASSIVE (Matrix of Average Sub-Subset

Information for Variable Elimination) a new method of feature selection and MRNET (Minimum

Redundancy NETwork), a new algorithm of network inference. Both tools rely on

the computation of mutual information, an information-theoretic measure of dependency.

More precisely, MASSIVE and MRNET use approximations of the mutual information

between a subset of variables and a target variable based on combinations of mutual informations

between sub-subsets of variables and the target. The used approximations allow

to estimate a series of low variate densities instead of one large multivariate density. Low

variate densities are well-suited for dealing with high variable-to-sample ratio datasets,

since they are rather cheap in terms of computational cost and they do not require a large

amount of samples in order to be estimated accurately. Numerous experimental results

show the competitiveness of these new approaches. Finally, our thesis has led to a freely

available source code of MASSIVE and an open-source R and Bioconductor package of

network inference.},
note = {Funder: Universite Libre de Bruxelles},
keywords = {},
pubstate = {published},
tppubtype = {phdthesis}
}