Extensions et applications des réseaux de neurones de graphes
Guillaume Lachaud · theses.fr (ABES) · 2023
Graphs are used everywhere to represent interactions between entities, whether physical such as atoms, molecules or people, or more abstract such as cities, friendships, ideas, etc. Amongst all the methods of machine learning that can be used, the recent advances in deep learning have made graph neural networks the de facto standard for graph representation learning. This thesis can be divided in two parts. First, we review the theoretical underpinnings of the most powerful graph neural networks. Second, we explore the challenges faced by the existing models when training on real world graph data. The powerfulness of a graph neural network is defined in terms of its expressiveness, i.e., its ability to distinguish non isomorphic graphs; or, in an equivalent manner, its ability to approximate permutation invariant and equivariant functions. We distinguish two broad families of the most powerful models. We summarise the mathematical properties as well as the advantages and disadvantages of these models in practical situations. Apart from the choice of the architecture, the quality of the graph data plays a crucial role in the ability to learn useful representations. Several challenges are faced by graph neural networks given the intrinsic nature of graph data. In contrast to typical machine learning methods that deal with tabular data, graph neural networks need to consider not only the features of the nodes but also the interconnectedness between them. Due to the connections between nodes, training neural networks on graphs can be done in two settings: in transductive learning, the model can have access to the test features in the training phase; in the inductive setting, the test data remains unseen. We study the differences in terms of performance between inductive and transductive learning for the node classification task. Additionally, the features that are fed to a model can be noisy or even missing. In this thesis we evaluate these challenges on real world datasets, and we propose a novel architecture to perform missing data imputation on graphs. Finally, while graphs can be the natural way to describe interactions, other types of data can benefit from being converted into graphs. In this thesis, we perform preliminary work on how to extract the most important parts of skin lesion images that could be used to create graphs and learn hidden relations in the data.