Embedding dynamic graphs
As we saw in Chapter 2, Graph Machine Learning, most of the state-of-the-art machine learning algorithms on graphs can be modeled into an encoder-decoder framework. The same applies to dynamic graphs. More specifically:
- The encoder takes as input a dynamic graph and returns as output its embedded representation
- The decoder takes as input an embedded representation of the dynamic graph and, depending on the task, outputs a prediction (it can be a new line, a class, or even a reconstructed graph)
In fields where dynamic graphs can be used to describe various phenomena, accurately modeling the graph’s evolution is often essential for precise predictions. Over time, various categories of machine learning models have been created to capture both the structure and evolution of dynamic graphs.
Notably, adaptations of graph neural networks (GNNs) tailored for dynamic graphs have recently demonstrated success in various domains...
