Saerens, MarcoLebichot, BertrandDe Plaen, Pierre-FrançoisPierre-FrançoisDe Plaen2025-05-142025-05-142025-05-142021https://hdl.handle.net/2078.2/23212Due to the rapid growth of the internet, there are more and more graph-structured datasets such as social networks, co-purchase graphs, citation networks, and so on. The combination of both the numerical features and the graph structure of these datasets can improve the classification abilities. In the last few years, convolutional graph Neural Networks have shown encouraging results for semi-supervised classification on graph-structured datasets. However, we show through a theoretical analysis and a series of experiments that part of these models are not able to correctly classify nodes when the graph structure does not meet the autocorrelation hypothesis and suggest this is due to the low-pass behavior of these models. Only the Graph Wavelet Neural Network successfully predicts the class of the graph nodes on both feature-driven and graph-driven datasets. We present a novel approach that further improves the model by working at different wavelet scales. We demonstrate in a number of experiments that our approach outperforms the one scale approach on challenging datasets by a significant margin. Our work also shows that model regularization does in general not improve the generalization capabilities of convolutional graph Neural Networks, despite being heavily used in practice. We exhibit, however, that graph based regularization might help. Additionally, we provide an open-source toolbox to help implement and compare classification models on graph-structured data. The toolbox is available at: https://github.com/pfdp0/graph-classification-TF2-toolbox.ClassificationGraphsSemi-supervisedConvolutionalSpectralSpatialWaveletMultiscaletext::thesis::master thesisthesis:30705