• cascade correlation
• deep learning
• graph
• machine learning

The application of the cascade correlation algorithm to automatically construct deep neural networks in order to implement classification or regression models for graph input data has already been proposed by A. Micheli with the "Neural Network for Graph" (NN4G) model. The aim of this thesis is to propose novel architectures as extensions of the NN4G model by evaluating new cascade expansion strategies, unit types, graph pooling methods, and a validation-based stop condition, and assessing their performance within a double cross-validation framework using nine popular graph classification tasks. Our models have achieved performances superior or comparable to state-of-the-art models that have already been evaluated on the same datasets with the same validation framework, are much more succinct (i.e. they require a considerably lower number of units, overall), and can be trained in the order of minutes. We have also discovered the crucial role played by layer width on the generalization ability of cascade models, thus allowing performances hitherto unattainable by the classic NN4G model.