Variants of GNNs
Several attempts have been made in recent years to address the problem of learning from graph data. Variants of the previously described GNN have been proposed, with the aim of improving its representation learning capability. Some of them are specifically designed to process specific types of graphs (direct, indirect, weighted, unweighted, static, dynamic, and so on).
Graph convolutional network (GCN)-based encoders are one of the most diffused variants of GNN for unsupervised learning. GCNs are GNN models inspired by many of the basic ideas behind CNNs. Filter parameters are typically shared over all locations in the graph and several layers are concatenated to form a deep network.
There are essentially two types of convolutional operations for graph data – namely, spectral approaches and non-spectral (spatial) approaches. The first, as the name suggests, defines convolution in the spectral domain (that is, decomposing graphs in a combination of simpler...
