DL on networks

In this section, we’ll consider a new type of DL model called GNNs, which process and operate on networks by embedding vertex, edge, or global properties of the network to learn outcomes related to individual networks, vertex properties within a network, or edge properties within a network. Essentially, the DL architecture evolves the topology of these embeddings to find key topological features in the input data that are predictive of the outcome. This can be done in a fully supervised or semi-supervised fashion. In this example, we’ll focus on SSL, where only some of the labels are known; however, by providing all labels as input, this can be changed to an SL setting.

Before we dive into the technical details of GNNs, let’s explore their use cases in more depth. Classifying networks themselves often yields important insight into problems such as image features or type, molecular compound toxicity or potential use as a pharmaceutical agent, or...