Kernelized Convolutional Operator for Graph Neural Networks

In recent years the very application on machine learning involve Neural Network model based on tensor, but there exists many fields of research that are suited with data structure based on graph, like proteins and chemistry compounds. In this context, the Graph Neural Network (GNN) models have been developed, they are tipically based on message passing which lose information about the structure since the convolution is defined in a non-structural way. The GNNs are still not fully explored in the machine learning field, for this reason in this thesis we present a generalized Graph Neaural Network model that use Kernel operator and Convolution on graphs in order to approach a graph classification problem. We want to apply the convolution using a graph as mask to be compared with the subgraphs of the main graph using a graph Kernel, in this way we maintain the data structure and we don’t lose information about it. The Goal of this project is to understand if with this approach the model can be trained and learn the patterns of the graph and produce relevant result.