Alippi, Cesare Livi, Lorenzo Grattarola, Daniele 2021-12-10 This thesis explores the field of graph neural networks, a class of deep learning models designed to learn representations of graphs. We organise the work into two parts. In the first part, we focus on the essential building blocks of graph neural networks. We present three novel operators for learning graph representations: one graph convolutional layer and two methods for pooling. We put particular emphasis on the topic of pooling, introducing a universal and modular framework to describe pooling operators, a taxonomy to organise the literature, and a set of evaluation criteria to assess an operator’s performance. The second part focuses on specific graph neural network architectures and their applications to cutting-edge problems in dynamical systems and computational biology. We present three main contributions. First, we introduce an autoencoder architecture for learning graph representations in non-Euclidean spaces. We apply our model to the tasks of molecule generation and change detection in graph sequences. Second, we propose a graph neural network designed to be interpretable, specifically to solve the problem of seizure localisation in subjects with epilepsy. Finally, we discuss the design of autoregressive models for sequences of graphs. application/pdf https://localhost:5000/ark:/12658/srd1319253 https://susi.usi.ch/global/documents/319253 https://susi.usi.ch/documents/319253/files/2021INFO014.pdf eng info:eu-repo/semantics/altIdentifier/urn/urn:nbn:ch:rero-006-119592 info:eu-repo/semantics/altIdentifier/ark/12658/srd1319253 info:eu-repo/semantics/openAccess License undefined Graph neural networks Geometric deep learning Deep learning Graphs Networks info:eu-repo/classification/udc/004 Graph neural networks : operators and architectures http://purl.org/coar/resource_type/c_db06