A review of deep learning on graphs

Deep learning on graphs The topic of this thesis is Deep learning on graphs. Deep learning is the subfield of machine learning studying algorithms and models constituted by several layers of functions learning in an increasingly level of abstraction the representation of a concept. “On graphs” refers to the application of this class of methodologies to data that are inherently represented as a graph structure, requiring ad hoc strategies and theoretical machinery to perform tasks of various nature over them. In particular, this work studies, from a compilative point of view, a relatively recent model of deep learning called Generative Adversarial Network (GAN). GANs are constituted by two neural networks, the most common form of deep learning, trying to beat each other in an optimization problem, in which one network, the generator, tries to deceive the other, the discriminator, producing examples that look similar to those belonging to a dataset of interest. At the end of the training of the networks, the generator should have learned the probability distribution of the dataset, allowing the algorithm to produce new, never seen, realistic examples. The first two chapters of this thesis are devoted to the description of these concepts and to the study of the best strategies allowing a smooth learning process. In the following two chapters, various architectures are described, referring to specific domains of application of GANs, with the main focus over the problems that arise when the previously explored concepts are applied to graph structured data, studying which are the currently available solutions in the literature.