Temporal Graph Neural Networks
In the previous chapters, we have only considered graphs where edges and features do not change. However, in the real world, there are many applications where this is not the case. For instance, in social networks, people follow and unfollow other users, posts go viral, and profiles evolve over time. This dynamicity cannot be represented using the GNN architectures we previously described. Instead, we must embed a new temporal dimension to transform static graphs into dynamic ones. These dynamic networks will then be used as inputs for a new family of GNNs: Temporal Graph Neural Networks (T-GNNs), also called Spatio-Temporal GNNs.
In this chapter, we will describe two kinds of dynamic graphs that include spatiotemporal information. We will list different applications and focus on time series forecasting, where temporal GNNs are mainly applied. The second section is dedicated to an application we previously looked at: web traffic forecasting. This time...
