Goal of this master's thesis was in cooperation with the company Avast to design a system, which can extract knowledge from a database of graphs. Graphs, used for data mining, describe behaviour of computer systems and they are anonymously inserted into the company's database from systems of the company's products users. Each graph in the database can be assigned with one of two labels: clean or malware (malicious) graph. The task of the proposed self-learning system is to find clusters of graphs in the graph database, in which the classes of graphs do not mix. Graph clusters with only one class of graphs can be interpreted as different types of clean or malware graphs and they are a useful source of further analysis on the graphs. To evaluate the quality of the clusters, a custom metric, named as monochromaticity, was designed. The metric evaluates the quality of the clusters based on how much clean and malware graphs are mixed in the clusters. The best results of the metric were obtained when vector representations of graphs were created by a deep learning model (variational  graph autoencoder with two relation graph convolution operators) and the parameterless method MeanShift was used for clustering over vectors.

In recent biomedical research, graph representation learning and cell profiling techniques have emerged as transformative tools for analyzing high-dimensional biological data. The integration of these methods, as investigated in this study, has facilitated an enhanced understanding of complex biological systems, consequently improving drug discovery. The research aimed to decipher connections between chemical structures and cellular phenotypes while incorporating other biological information like proteins and pathways into the workflow. To achieve this, machine learning models' efficacy was examined for classification and regression tasks. The newly proposed graph-level and bio-graph integrative predictors were compared with traditional models. Results demonstrated their potential, particularly in classification tasks. Moreover, the topology of the COVID-19 BioGraph was analyzed, revealing the complex interconnections between chemicals, proteins, and biological pathways. By combining network analysis, graph representation learning, and statistical methods, the study was able to predict active chemical combinations within inactive compounds, thereby exhibiting significant potential for further investigations. Graph-based generative models were also used for molecule generation opening up further research avenues in finding lead compounds. In conclusion, this study underlines the potential of combining graph representation learning and cell profiling techniques in advancing biomedical research in drug repurposing and drug combination. This integration provides a better understanding of complex biological systems, assists in identifying therapeutic targets, and contributes to optimizing molecule generation for drug discovery. Future investigations should optimize these models and validate the drug combination discovery approach. As these techniques continue to evolve, they hold the potential to significantly impact the future of drug screening, drug repurposing, and drug combinations.

En informatique, la résolution de problèmes NP-difficiles en un temps raisonnable est d’une grande importance : optimisation de la chaîne d’approvisionnement, planification, routage, alignement de séquences biologiques multiples, inference dans les modèles graphiques pro- babilistes, et même certains problèmes de cryptographie sont tous des examples de la classe NP-complet. En pratique, nous modélisons beaucoup d’entre eux comme un problème d’op- timisation en nombre entier, que nous résolvons à l’aide de la méthodologie séparation et évaluation. Un algorithme de ce style divise un espace de recherche pour l’explorer récursi- vement (séparation), et obtient des bornes d’optimalité en résolvant des relaxations linéaires sur les sous-espaces (évaluation). Pour spécifier un algorithme, il faut définir plusieurs pa- ramètres, tel que la manière d’explorer les espaces de recherche, de diviser une recherche l’espace une fois exploré, ou de renforcer les relaxations linéaires. Ces politiques peuvent influencer considérablement la performance de résolution. Ce travail se concentre sur une nouvelle manière de dériver politique de recherche, c’est à dire le choix du prochain sous-espace à séparer étant donné une partition en cours, en nous servant de l’apprentissage automatique profond. Premièrement, nous collectons des données résumant, sur une collection de problèmes donnés, quels sous-espaces contiennent l’optimum et quels ne le contiennent pas. En représentant ces sous-espaces sous forme de graphes bipartis qui capturent leurs caractéristiques, nous entraînons un réseau de neurones graphiques à déterminer la probabilité qu’un sous-espace contienne la solution optimale par apprentissage supervisé. Le choix d’un tel modèle est particulièrement utile car il peut s’adapter à des problèmes de différente taille sans modifications. Nous montrons que notre approche bat celle de nos concurrents, consistant à des modèles d’apprentissage automatique plus simples entraînés à partir des statistiques du solveur, ainsi que la politique par défaut de SCIP, un solveur open-source compétitif, sur trois familles NP-dures: des problèmes de recherche de stables de taille maximum, de flots de réseau multicommodité à charge fixe, et de satisfiabilité maximum. / In computer science, solving NP-hard problems in a reasonable time is of great importance, such as in supply chain optimization, scheduling, routing, multiple biological sequence align- ment, inference in probabilistic graphical models, and even some problems in cryptography. In practice, we model many of them as a mixed integer linear optimization problem, which we solve using the branch and bound framework. An algorithm of this style divides a search space to explore it recursively (branch) and obtains optimality bounds by solving linear relaxations in such sub-spaces (bound). To specify an algorithm, one must set several pa- rameters, such as how to explore search spaces, how to divide a search space once it has been explored, or how to tighten these linear relaxations. These policies can significantly influence resolution performance. This work focuses on a novel method for deriving a search policy, that is, a rule for select- ing the next sub-space to explore given a current partitioning, using deep machine learning. First, we collect data summarizing which subspaces contain the optimum, and which do not. By representing these sub-spaces as bipartite graphs encoding their characteristics, we train a graph neural network to determine the probability that a subspace contains the optimal so- lution by supervised learning. The choice of such design is particularly useful as the machine learning model can automatically adapt to problems of different sizes without modifications. We show that our approach beats the one of our competitors, consisting of simpler machine learning models trained from solver statistics, as well as the default policy of SCIP, a state- of-the-art open-source solver, on three NP-hard benchmarks: generalized independent set, fixed-charge multicommodity network flow, and maximum satisfiability problems.

[pt] Nesta tese, propomos novas arquiteturas para redes neurais profundas utlizando métodos de atenção e álgebra multilinear para aumentar seu desempenho. Também exploramos convoluções em grafos e suas particularidades. Nos concentramos aqui em problemas relacionados à estimativa de pose em tempo real. A estimativa de pose é um problema desafiador em visão computacional com muitas aplicações reais em áreas como realidade aumentada, realidade virtual, animação por computador e reconstrução de cenas 3D. Normalmente, o problema a ser abordado envolve estimar a pose humana 2D ou 3D, ou seja, as partes do corpo de pessoas em imagens ou vídeos, bem como seu posicionamento e estrutura. Diveros trabalhos buscam atingir alta precisão usando arquiteturas baseadas em redes neurais de convolução convencionais; no entanto, erros causados por oclusão e motion blur não são incomuns, e ainda esses modelos são computacionalmente pesados para aplicações em tempo real. Exploramos diferentes arquiteturas para melhorar o tempo de processamento destas redes e, como resultado, propomos dois novos modelos de rede neural para estimativa de pose 2D e 3D. Também apresentamos uma nova arquitetura para redes de atenção em grafos chamada de atenção em grafos semânticos. / [en] This thesis proposes new architectures for deep neural networks with attention enhancement and multilinear algebra methods to increase their performance. We also explore graph convolutions and their particularities. We focus here on the problems related to real-time pose estimation. Pose estimation is a challenging problem in computer vision with many real applications in areas including augmented reality, virtual reality, computer animation, and 3D scene reconstruction. Usually, the problem to be addressed involves estimating the 2D and 3D human pose, i.e., the anatomical keypoints or body parts of persons in images or videos. Several papers propose approaches to achieve high accuracy using architectures based on conventional convolution neural networks; however, mistakes caused by occlusion and motion blur are not uncommon, and those models are computationally very intensive for real-time applications. We explore different architectures to improve processing time, and, as a result, we propose two novel neural network models for 2D and 3D pose estimation. We also introduce a new architecture for Graph attention networks called Semantic Graph Attention.

<p>Artificial neural networks (ANNs) have achieved extraordinary performance in various domains in recent years. However, some studies reveal that ANNs may be vulnerable in three aspects: label scarcity, perturbations, and open-set emerging classes. Noisy labeling and self-supervised learning approaches address the label scarcity issues, but most of the work couldn't handle the perturbations. Adversarial training methods, topological denoising methods, and mechanism designing methods aim to mitigate the negative effects caused by perturbations. However, adversarial training methods can barely train a robust model under the circumstance of extensive label scarcity; topological denoising methods are not efficient on dynamic data structures; and mechanism designing methods often depend on heuristic explorations. Detection-based methods devote to identifying novel or anomaly instances for further downstream tasks. Nonetheless, such instances may belong to open-set new emerging classes. To embrace the aforementioned challenges, we address the robustness issues of ANNs from two aspects. First, we propose a series of Bayesian label transition models to improve the robustness of Graph Neural Networks (GNNs) in the presence of label scarcity and perturbations in the graph domain. Second, we propose a new non-exhaustive learning model, named NE-GM-GAN, to handle both open-set problems and class-imbalance issues in network intrusion datasets. Extensive experiments with several datasets demonstrate that our proposed models can effectively improve the robustness of ANNs.</p>

ICT systems are part of the vital infrastructure in today’s society. These systems are under constant threat and efforts are continually being put forth by cyber security experts to protect them. By applying modern AI methods, can these efforts both be improved and alleviated of the cost of expert work. This thesis examines whether a reinforcement learning (RL) algorithm can be applied to a cyber security modelling of ICT systems. The research question answered is that of how well an RL algorithm can optimise the resource cost of successful cyber attacks, as represented by a cyber security model? The modelling, called Meta Attack Language (MAL), is a meta language for attack graphs that details the individual steps to be taken in a cyber attack. In the previous work of Manuel Rickli’s thesis, a method of automatically generating attack graphs according to MAL aimed at modelling industry-level computer networks, was presented. The method was used to generate different distributions of attack graphs that were used to train deep Q-learning (DQN) agents. The agents’ results were then compared with a random agent and a greedy method based on the A∗ search algorithm. The results show that attack step selection can be achieved with a higher performance than the uninformed choice of the random agent, by DQN. However, DQN was unable to achieve higher performance than the A∗ method. This may be due to the simplicity of the attack graph generation or the fact that the A∗ method has access to the complete attack graph, amongst other factors. The thesis also raises questions about general representation of MAL attack graphs as RL problems and how to apply RL algorithms to the RL problem. The source code of this thesis is available at: https://github.com/KTH-SSAS/sandor-berglund-thesis. / IT-system är i dagens samhälle en väsentlig del av infrastrukturen som är under konstant hot av olika personer och organisationer. IT-säkerhetsexperter lägger ner beständigt arbete på att hålla dessa system säkra och för att avvärja illvilliga auktioner mot IT-system. Moderna AI-metoder kan användas för att förbättra och lätta på kostnaden av expertarbetet inom området. Detta examensarbete avser att undersöka hur en förstärkningsinlärningsalgoritm kan appliceras på en cybersäkerhetsmodell. Det görs genom att besvara frågeställningen: Hur väl kan en förstärkningsinlärningsalgoritm optimera en cyberattack representerat av en cybersäkerhetsmodell? Meta Attack Language (MAL) är ett metaspråk för attackgrafer som beskriver varje steg i en cyberattack. I detta examensarbete användes Manuell Ricklis implementation av MAL samt attack grafs generation för att definiera ett förstärkningsinlärningsproblem. Förstärkningsinlärningsalgoritmen deep Q-learning (DQN) användes för att träna ett attention baserat neuronnät på olika fördelningar av attackgrafer och jämfördes med en slumpmässig agent och en girig metod baserad på sökalgoritmen A∗ . Resultaten visar att DQN kunde producera en agent som presterar bättre än den oinformerade slumpmässiga agenten. Agenten presterade däremot inte bättre än den giriga A∗ metoden, vilket kan bero på att A∗ har tillgång till den fulla attack grafen, bland andra bidragande faktorer. Arbetet som läggs fram här väcker frågor om hur MAL-attackgrafer representeras som förstärkningsinlärningsproblem och hur förstärkningsinlärningsalgoritmer appliceras där av. Källkoden till det här examensarbetet finns på: https://github.com/KTHSSAS/sandor-berglund-thesis.

This work presents an investigation of tailoring Network Representation Learning (NRL) for an application in the Financial Industry. NRL approaches are data-driven models that learn how to encode graph structures into low-dimensional vector spaces, which can be further exploited by downstream Machine Learning applications. They can potentially bring a lot of benefits in the Financial Industry since they extract in an automatic way features that can provide useful input regarding graph structures, called embeddings. Financial transactions can be represented as a network, and through NRL, it is possible to extract embeddings that reflect the intrinsic inter-connected nature of economic relationships. Such embeddings can be used for several purposes, among which Anomaly Detection to fight financial crime.This work provides a qualitative analysis over state-of-the-art NRL models, which identifies Graph Convolutional Network (ConvGNN) as the most suitable category of approaches for Financial Industry but with a certain need for further improvement. Financial Industry poses additional challenges when modelling a NRL solution. Despite the need of having a scalable solution to handle real-world graph with considerable dimensions, it is necessary to take into consideration several characteristics: transactions graphs are inherently dynamic since every day new transactions are executed and nodes can be heterogeneous. Besides, everything is further complicated by the need to have updated information in (near) real-time due to the sensitivity of the application domain. For these reasons, GraphSAGE has been considered as a base for the experiments, which is an inductive ConvGNN model. Two variants of GraphSAGE are presented: a dynamic variant whose weights evolve accordingly with the input sequence of graph snapshots, and a variant specifically meant to handle bipartite graphs. These variants have been evaluated by applying them to real-world data and leveraging the generated embeddings to perform Anomaly Detection. The experiments demonstrate that leveraging these variants leads toimagecomparable results with other state-of-the-art approaches, but having the advantage of being suitable to handle real-world financial data sets. / Detta arbete presenterar en undersökning av tillämpningar av Network Representation Learning (NRL) inom den finansiella industrin. Metoder inom NRL möjliggör datadriven kondensering av grafstrukturer till lågdimensionella och lätthanterliga vektorer.Dessa vektorer kan sedan användas i andra maskininlärningsuppgifter. Närmare bestämt, kan metoder inom NRL underlätta hantering av och informantionsutvinning ur beräkningsintensiva och storskaliga grafer inom den finansiella sektorn, till exempel avvikelsehantering bland finansiella transaktioner. Arbetet med data av denna typ försvåras av det faktum att transaktionsgrafer är dynamiska och i konstant förändring. Utöver detta kan noderna, dvs transaktionspunkterna, vara vitt skilda eller med andra ord härstamma från olika fördelningar.I detta arbete har Graph Convolutional Network (ConvGNN) ansetts till den mest lämpliga lösningen för nämnda tillämpningar riktade mot upptäckt av avvikelser i transaktioner. GraphSAGE har använts som utgångspunkt för experimenten i två olika varianter: en dynamisk version där vikterna uppdateras allteftersom nya transaktionssekvenser matas in, och en variant avsedd särskilt för bipartita (tvådelade) grafer. Dessa varianter har utvärderats genom användning av faktiska datamängder med avvikelsehantering som slutmål.

Tesis por compendio / [ES] La detección del lenguaje abusivo es una tarea que se ha vuelto cada vez más importante en la era digital moderna, donde la comunicación se produce a través de diversas plataformas en línea. El aumento de las interacciones en estas plataformas ha provocado un aumento de la aparición del lenguaje abusivo. Abordar dicho contenido es crucial para mantener un entorno en línea seguro e inclusivo. Sin embargo, esta tarea enfrenta varios desafíos que la convierten en un área compleja y que demanda de continua investigación y desarrollo. En particular, detectar lenguaje abusivo en entornos con escasez de datos presenta desafíos adicionales debido a que el desarrollo de sistemas automáticos precisos a menudo requiere de grandes conjuntos de datos anotados. En esta tesis investigamos diferentes aspectos de la detección del lenguaje abusivo, prestando especial atención a entornos con datos limitados. Primero, estudiamos el sesgo hacia palabras clave abusivas en modelos entrenados para la detección del lenguaje abusivo. Con este propósito, proponemos dos métodos para extraer palabras clave potencialmente abusivas de colecciones de textos. Luego evaluamos el sesgo hacia las palabras clave extraídas y cómo se puede modificar este sesgo para influir en el rendimiento de la detección del lenguaje abusivo. El análisis y las conclusiones de este trabajo revelan evidencia de que es posible mitigar el sesgo y que dicha reducción puede afectar positivamente el desempeño de los modelos. Sin embargo, notamos que no es posible establecer una correspondencia similar entre la variación del sesgo y el desempeño de los modelos cuando hay escasez datos con las técnicas de reducción del sesgo estudiadas. En segundo lugar, investigamos el uso de redes neuronales basadas en grafos para detectar lenguaje abusivo. Por un lado, proponemos una estrategia de representación de textos diseñada con el objetivo de obtener un espacio de representación en el que los textos abusivos puedan distinguirse fácilmente de otros textos. Por otro lado, evaluamos la capacidad de redes neuronales convolucionales basadas en grafos para clasificar textos abusivos. La siguiente parte de nuestra investigación se centra en analizar cómo el aumento de datos puede influir en el rendimiento de la detección del lenguaje abusivo. Para ello, investigamos dos técnicas bien conocidas basadas en el principio de minimización del riesgo en la vecindad de instancias originales y proponemos una variante para una de ellas. Además, evaluamos técnicas simples basadas en el reemplazo de sinónimos, inserción aleatoria, intercambio aleatorio y eliminación aleatoria de palabras. Las contribuciones de esta tesis ponen de manifiesto el potencial de las redes neuronales basadas en grafos y de las técnicas de aumento de datos para mejorar la detección del lenguaje abusivo, especialmente cuando hay limitación de datos. Estas contribuciones han sido publicadas en conferencias y revistas internacionales. / [CA] La detecció del llenguatge abusiu és una tasca que s'ha tornat cada vegada més important en l'era digital moderna, on la comunicació es produïx a través de diverses plataformes en línia. L'augment de les interaccions en estes plataformes ha provocat un augment de l'aparició de llenguatge abusiu. Abordar este contingut és crucial per a mantindre un entorn en línia segur i inclusiu. No obstant això, esta tasca enfronta diversos desafiaments que la convertixen en una àrea complexa i contínua de recerca i desenvolupament. En particular, detectar llenguatge abusiu en entorns amb escassetat de dades presenta desafiaments addicionals pel fet que el desenvolupament de sistemes automàtics precisos sovint requerix de grans conjunts de dades anotades. En esta tesi investiguem diferents aspectes de la detecció del llenguatge abusiu, prestant especial atenció a entorns amb dades limitades. Primer, estudiem el biaix cap a paraules clau abusives en models entrenats per a la detecció de llenguatge abusiu. Amb este propòsit, proposem dos mètodes per a extraure paraules clau potencialment abusives de col·leccions de textos. Després avaluem el biaix cap a les paraules clau extretes i com es pot modificar este biaix per a influir en el rendiment de la detecció de llenguatge abusiu. L'anàlisi i les conclusions d'este treball revelen evidència que és possible mitigar el biaix i que esta reducció pot afectar positivament l'acompliment dels models. No obstant això, notem que no és possible establir una correspondència similar entre la variació del biaix i l'acompliment dels models quan hi ha escassetat dades amb les tècniques de reducció del biaix estudiades. En segon lloc, investiguem l'ús de xarxes neuronals basades en grafs per a detectar llenguatge abusiu. D'una banda, proposem una estratègia de representació textual dissenyada amb l'objectiu d'obtindre un espai de representació en el qual els textos abusius puguen distingir-se fàcilment d'altres textos. D'altra banda, avaluem la capacitat de models basats en xarxes neuronals convolucionals basades en grafs per a classificar textos abusius. La següent part de la nostra investigació se centra en analitzar com l'augment de dades pot influir en el rendiment de la detecció del llenguatge abusiu. Per a això, investiguem dues tècniques ben conegudes basades en el principi de minimització del risc en el veïnatge d'instàncies originals i proposem una variant per a una d'elles. A més, avaluem tècniques simples basades en el reemplaçament de sinònims, inserció aleatòria, intercanvi aleatori i eliminació aleatòria de paraules. Les contribucions d'esta tesi destaquen el potencial de les xarxes neuronals basades en grafs i de les tècniques d'augment de dades per a millorar la detecció del llenguatge abusiu, especialment quan hi ha limitació de dades. Estes contribucions han sigut publicades en revistes i conferències internacionals. / [EN] Abusive language detection is a task that has become increasingly important in the modern digital age, where communication takes place via various online platforms. The increase in online interactions has led to an increase in the occurrence of abusive language. Addressing such content is crucial to maintaining a safe and inclusive online environment. However, this task faces several challenges that make it a complex and ongoing area of research and development. In particular, detecting abusive language in environments with sparse data poses an additional challenge, since the development of accurate automated systems often requires large annotated datasets. In this thesis we investigate different aspects of abusive language detection, paying particular attention to environments with limited data. First, we study the bias toward abusive keywords in models trained for abusive language detection. To this end, we propose two methods for extracting potentially abusive keywords from datasets. We then evaluate the bias toward the extracted keywords and how this bias can be modified in order to influence abusive language detection performance. The analysis and conclusions of this work reveal evidence that it is possible to mitigate the bias and that such a reduction can positively affect the performance of the models. However, we notice that it is not possible to establish a similar correspondence between bias mitigation and model performance in low-resource settings with the studied bias mitigation techniques. Second, we investigate the use of models based on graph neural networks to detect abusive language. On the one hand, we propose a text representation framework designed with the aim of obtaining a representation space in which abusive texts can be easily distinguished from other texts. On the other hand, we evaluate the ability of models based on convolutional graph neural networks to classify abusive texts. The next part of our research focuses on analyzing how data augmentation can influence the performance of abusive language detection. To this end, we investigate two well-known techniques based on the principle of vicinal risk minimization and propose a variant for one of them. In addition, we evaluate simple techniques based on the operations of synonym replacement, random insertion, random swap, and random deletion. The contributions of this thesis highlight the potential of models based on graph neural networks and data augmentation techniques to improve abusive language detection, especially in low-resource settings. These contributions have been published in several international conferences and journals. / This research work was partially funded by the Spanish Ministry of Science and Innovation under the research project MISMIS-FAKEnHATE on Misinformation and Miscommunication in social media: FAKE news and HATE speech (PGC2018-096212-B-C31). The authors thank also the EU-FEDER Comunitat Valenciana 2014-2020 grant IDIFEDER/2018/025. This work was done in the framework of the research project on Fairness and Transparency for equitable NLP applications in social media, funded by MCIN/AEI/10.13039/501100011033 and by ERDF, EU A way of making EuropePI. FairTransNLP research project (PID2021-124361OB-C31) funded by MCIN/AEI/10.13039/501100011033 and by ERDF, EU A way of making Europe. Part of the work presented in this article was performed during the first author’s research visit to the University of Mannheim, supported through a Contact Fellowship awarded by the DAAD scholarship program “STIBET Doktoranden”. / Peña Sarracén, GLDL. (2024). On the Keyword Extraction and Bias Analysis, Graph-based Exploration and Data Augmentation for Abusive Language Detection in Low-Resource Settings [Tesis doctoral]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/203266 / Compendio

Humans' incredible capacity to model the complexity of the physical world is possible because they cast this complexity as the composition of simpler entities and rules to process them. Extensive work in cognitive science indeed shows that human perception and reasoning ability is structured around objects. Motivated by this observation, a growing number of recent work focused on entity-centric approaches to learning representation and their potential to facilitate downstream tasks. In the first contribution, we show how an entity-centric approach to learning a transition model allows us to extract meaningful visual entities and to learn transition rules that achieve better compositional generalization. In the second contribution, we show how an entity-centric approach to generating graphs allows us to design a model for conditional graph generation that permits direct optimisation of the graph properties. We investigate the performance of our model in a prototype-based molecular graph generation task. In this task, called lead optimization in drug discovery, we wish to adjust a few physico-chemical properties of a molecule that has proven efficient in vitro in order to make a drug out of it. / L'incroyable capacité des humains à modéliser la complexité du monde physique est rendue possible par la décomposition qu'ils en font en un ensemble d'entités et de règles simples. De nombreux travaux en sciences cognitives montre que la perception humaine et sa capacité à raisonner est essentiellement centrée sur la notion d'objet. Motivés par cette observation, de récents travaux se sont intéressés aux différentes approches d'apprentissage de représentations centrées sur des entités et comment ces représentations peuvent être utilisées pour résoudre plus facilement des tâches sous-jacentes. Dans la première contribution on montre comment une architecture centrée sur la notion d'entité va permettre d'extraire des entités visuelles interpretables et d'apprendre un modèle du monde plus robuste aux différentes configurations d'objets. Dans la deuxième contribution on s’intéresse à un modèle de génération de graphes dont l'architecture est également centrée sur la notion d'entités et comment cette architecture rend plus facile l'apprentissage d'une génération conditionelle à certaines propriétés du graphe. On s’intéresse plus particulièrement aux applications en découverte de médicaments. Dans cette tâche, on souhaite optimiser certaines propriétés physico-chmiques du graphe d'une molécule qui a été efficace in-vitro et dont on veut faire un médicament.

Évaluer la vulnérabilité des réseaux est un enjeu de plus en plus critique. Dans ce mémoire, nous nous penchons sur une approche étudiant la défense d’infrastructures stratégiques contre des attaques malveillantes au travers de problèmes d'optimisations multiniveaux. Plus particulièrement, nous analysons un jeu séquentiel en trois étapes appelé le « Multilevel Critical Node problem » (MCN). Ce jeu voit deux joueurs s'opposer sur un graphe: un attaquant et un défenseur. Le défenseur commence par empêcher préventivement que certains nœuds soient attaqués durant une phase de vaccination. Ensuite, l’attaquant infecte un sous ensemble des nœuds non vaccinés. Finalement, le défenseur réagit avec une stratégie de protection. Dans ce mémoire, nous fournissons les premiers résultats de complexité pour MCN ainsi que ceux de ses sous-jeux. De plus, en considérant les différents cas de graphes unitaires, pondérés ou orientés, nous clarifions la manière dont la complexité de ces problèmes varie. Nos résultats contribuent à élargir les familles de problèmes connus pour être complets pour les classes NP, $\Sigma_2^p$ et $\Sigma_3^p$. Motivés par l’insolubilité intrinsèque de MCN, nous concevons ensuite une heuristique efficace pour le jeu. Nous nous appuyons sur les approches récentes cherchant à apprendre des heuristiques pour des problèmes d’optimisation combinatoire en utilisant l’apprentissage par renforcement et les réseaux de neurones graphiques. Contrairement aux précédents travaux, nous nous intéressons aux situations dans lesquelles de multiples joueurs prennent des décisions de manière séquentielle. En les inscrivant au sein du formalisme d’apprentissage multiagent, nous concevons un algorithme apprenant à résoudre des problèmes d’optimisation combinatoire multiniveaux budgétés opposant deux joueurs dans un jeu à somme nulle sur un graphe. Notre méthode est basée sur un simple curriculum : si un agent sait estimer la valeur d’une instance du problème ayant un budget au plus B, alors résoudre une instance avec budget B+1 peut être fait en temps polynomial quelque soit la direction d’optimisation en regardant la valeur de tous les prochains états possibles. Ainsi, dans une approche ascendante, nous entraînons notre agent sur des jeux de données d’instances résolues heuristiquement avec des budgets de plus en plus grands. Nous rapportons des résultats quasi optimaux sur des graphes de tailles au plus 100 et un temps de résolution divisé par 185 en moyenne comparé au meilleur solutionneur exact pour le MCN. / Evaluating the vulnerability of networks is a problem which has gain momentum in recent decades. In this work, we focus on a Multilevel Programming approach to study the defense of critical infrastructures against malicious attacks. We analyze a three-stage sequential game played in a graph called the Multilevel Critical Node problem (MCN). This game sees two players competing with each other: a defender and an attacker. The defender starts by preventively interdicting nodes from being attacked during what is called a vaccination phase. Then, the attacker infects a subset of non-vaccinated nodes and, finally, the defender reacts with a protection strategy. We provide the first computational complexity results associated with MCN and its subgames. Moreover, by considering unitary, weighted, undirected and directed graphs, we clarify how the theoretical tractability or intractability of those problems vary. Our findings contribute with new NP-complete, $\Sigma_2^p$-complete and $\Sigma_3^p$-complete problems. Motivated by the intrinsic intractability of the MCN, we then design efficient heuristics for the game by building upon the recent approaches seeking to learn heuristics for combinatorial optimization problems through graph neural networks and reinforcement learning. But contrary to previous work, we tackle situations with multiple players taking decisions sequentially. By framing them in a multi-agent reinforcement learning setting, we devise a value-based method to learn to solve multilevel budgeted combinatorial problems involving two players in a zero-sum game over a graph. Our framework is based on a simple curriculum: if an agent knows how to estimate the value of instances with budgets up to B, then solving instances with budget B+1 can be done in polynomial time regardless of the direction of the optimization by checking the value of every possible afterstate. Thus, in a bottom-up approach, we generate datasets of heuristically solved instances with increasingly larger budgets to train our agent. We report results close to optimality on graphs up to 100 nodes and a 185 x speedup on average compared to the quickest exact solver known for the MCN.