Graph neural networks for classification: models and applications

Graph neural networks have emerged in the past years as very promising methods for the analysis of graph-structured data. Useful insights can in fact be extracted by allowing learning models to take into account relationships between entities in a graph. The main methods used in the context of graph neural networks are here described and compared, with a focus on the extension of convolutional layers to graph structures. Afterwards, an analysis of how attention mechanisms integrate with graph neural networks is introduced. In this context a new method is proposed for allowing a graph neural network to attend over its own input in the context of graph classification. An application of these methods to biomedical data is finally presented, with an example in the field of Parkinson's disease classification.