Exploring GNNs

Using structured data as graphs in an ML model is problematic due to dimensionality and non-Euclidean properties. Researchers have tried to train ML models on graph-structured data by summarizing or representing information in a simplified way. But, this feels more like preprocessing than a real training process. GNNs help us create an end-to-end ML model trained to learn a representation of structured data in graphs and to fit a predictive model into it. In order to understand the operating principle of these algorithms, it is necessary to start from the basic concepts of graph theory.

Introducing graph theory

Graphs are rational mathematical structures that are used in various fields of study including mathematics, physics, and computer science up to topology, chemistry, and engineering. A graph is represented graphically by a structure of vertices and edges. The vertices can be seen as events from which different alternatives (the edge) start. Typically, graphics...