<p>Research on graph representation learning (GRL) has made major strides over the past decade, with widespread applications in domains such as e-commerce, personalization, fraud & abuse, life sciences, and social network analysis. Despite its widespread success, fundamental questions on practices employed in modern day GRL have remained unanswered. Unraveling and advancing two such fundamental questions on the practices in modern day GRL forms the overarching theme of my thesis.</p> <p>The first part of my thesis deals with the mathematical foundations of GRL. GRL is used to solve tasks such as node classification, link prediction, clustering, graph classification, and so on, albeit with seemingly different frameworks (e.g. Graph neural networks for node/graph classification, (implicit) matrix factorization for link prediction/ clustering, etc.). The existence of very distinct frameworks for different graph tasks has puzzled researchers and practitioners alike. In my thesis, using group theory, I provide a theoretical blueprint that connects these seemingly different frameworks, bridging methods like matrix factorization and graph neural networks. With this renewed understanding, I then provide guidelines to better realize the full capabilities of these methods in a multitude of tasks.</p> <p>The second part of my thesis deals with cases where modeling real-world objects as a graph is an oversimplified description of the underlying data. Specifically, I look at two such objects (i) modeling hypergraphs (where edges encompass two or more vertices) and (ii) using GRL for predicting protein properties. Towards (i) hypergraphs, I develop a hypergraph neural network which takes advantage of the inherent sparsity of real world hypergraphs, without unduly sacrificing on its ability to distinguish non isomorphic hypergraphs. The designed hypergraph neural network is then leveraged to learn expressive representations of hyperedges for two tasks, namely hyperedge classification and hyperedge expansion. Experiments show that using our network results in improved performance over the current approach of converting the hypergraph into a dyadic graph and using (dyadic) GRL frameworks. Towards (ii) proteins, I introduce the concept of conditional invariances and leverage it to model the inherent flexibility present in proteins. Using conditional invariances, I provide a new framework for GRL which can capture protein-dependent conformations and ensures that all viable conformers of a protein obtain the same representation. Experiments show that endowing existing GRL models with my framework shows noticeable improvements on multiple different protein datasets and tasks.</p>

The core research of this thesis, mostly comprising chapter four, has been accepted to the Learning on Graphs (LoG) 2022 conference for a spotlight presentation as a standalone paper, under the title "Taxonomy of Benchmarks in Graph Representation Learning", and is to be published in the Proceedings of Machine Learning Research (PMLR) series. As a main author of the paper, my specific contributions to this paper cover problem formulation, design and implementation of our taxonomy framework and experimental pipeline, collation of our results and of course the writing of the article. / L'apprentissage profond sur les graphes a atteint des niveaux de succès sans précédent ces dernières années grâce aux réseaux de neurones de graphes (GNN), des architectures de réseaux de neurones spécialisées qui ont sans équivoque surpassé les approches antérieurs d'apprentissage définies sur des graphes. Les GNN étendent le succès des réseaux de neurones aux données structurées en graphes en tenant compte de leur géométrie intrinsèque. Bien que des recherches approfondies aient été effectuées sur le développement de GNN avec des performances supérieures à celles des modèles références d'apprentissage de représentation graphique, les procédures d'analyse comparative actuelles sont insuffisantes pour fournir des évaluations justes et efficaces des modèles GNN. Le problème peut-être le plus répandu et en même temps le moins compris en ce qui concerne l'analyse comparative des graphiques est la "couverture de domaine": malgré le nombre croissant d'ensembles de données graphiques disponibles, la plupart d'entre eux ne fournissent pas d'informations supplémentaires et au contraire renforcent les biais potentiellement nuisibles dans le développement d’un modèle GNN. Ce problème provient d'un manque de compréhension en ce qui concerne les aspects d'un modèle donné qui sont sondés par les ensembles de données de graphes. Par exemple, dans quelle mesure testent-ils la capacité d'un modèle à tirer parti de la structure du graphe par rapport aux fonctionnalités des nœuds? Ici, nous développons une approche fondée sur des principes pour taxonomiser les ensembles de données d'analyse comparative selon un "profil de sensibilité" qui est basé sur la quantité de changement de performance du GNN en raison d'une collection de perturbations graphiques. Notre analyse basée sur les données permet de mieux comprendre quelles caractéristiques des données de référence sont exploitées par les GNN. Par conséquent, notre taxonomie peut aider à la sélection et au développement de repères graphiques adéquats et à une évaluation mieux informée des futures méthodes GNN. Enfin, notre approche et notre implémentation dans le package GTaxoGym (https://github.com/G-Taxonomy-Workgroup/GTaxoGym) sont extensibles à plusieurs types de tâches de prédiction de graphes et à des futurs ensembles de données. / Deep learning on graphs has attained unprecedented levels of success in recent years thanks to Graph Neural Networks (GNNs), specialized neural network architectures that have unequivocally surpassed prior graph learning approaches. GNNs extend the success of neural networks to graph-structured data by accounting for their intrinsic geometry. While extensive research has been done on developing GNNs with superior performance according to a collection of graph representation learning benchmarks, current benchmarking procedures are insufficient to provide fair and effective evaluations of GNN models. Perhaps the most prevalent and at the same time least understood problem with respect to graph benchmarking is "domain coverage": Despite the growing number of available graph datasets, most of them do not provide additional insights and on the contrary reinforce potentially harmful biases in GNN model development. This problem stems from a lack of understanding with respect to what aspects of a given model are probed by graph datasets. For example, to what extent do they test the ability of a model to leverage graph structure vs. node features? Here, we develop a principled approach to taxonomize benchmarking datasets according to a "sensitivity profile" that is based on how much GNN performance changes due to a collection of graph perturbations. Our data-driven analysis provides a deeper understanding of which benchmarking data characteristics are leveraged by GNNs. Consequently, our taxonomy can aid in selection and development of adequate graph benchmarks, and better informed evaluation of future GNN methods. Finally, our approach and implementation in the GTaxoGym package (https://github.com/G-Taxonomy-Workgroup/GTaxoGym) are extendable to multiple graph prediction task types and future datasets.