1. Software version
matlab2021a
and the Graph Neural Network Toolbox
2. Theoretical knowledge of this algorithm
In some scientific fields, such as machine vision, molecular chemistry, molecular biology, pattern recognition, and data acquisition, their data have underlying relationships that can be represented as graphs. In this paper, we postulate a new neural network model, called a graph neural network model (GNN); it extends the data processing capabilities of existing neural networks to the domain of graphs. This GNN model can directly process most graphs used in practice, such as cyclic graphs, acyclic graphs, directed graphs and undirected graphs, and use a function to map graph G and one of its nodes to m-dimensional Euclidean space. We derive a supervised learning algorithm to judge the parameters of this putative GNN model. The computational cost of the algorithm is also taken into consideration. We will verify the effect of the algorithm through experimental data and demonstrate its generalization ability.