Graph neural networks

DL algorithms make use of graphs to predict at the level of nodes, edges, or entire graphs. In node classification, the label of samples (nodes) is determined by looking at the labels of neighbors. In graph classification, the entire graph is classified into different categories, an example being categorizing documents using natural language processing. The relationships (edges) between nodes or entities are utilized in recommendation systems. Image and text are types of structured data that can be described as grids of pixels and sequences of words, respectively. These are shown in Figure 6.10a. Graphs, in contrast, are unstructured data. Graphs can contain any kind of data, including images and text.

Figure 6.10a: Structured data (L) as opposed to graphs/networks (R)

GNNs organize graphs using a process called message passing so that DL algorithms can use the embedded information about the neighbors of each node to find patterns...