Scaling Graph Learning for the Enterprise

Chapter 5. Graph Neural Networks

In this chapter, we will explore graph neural networks (GNNs), an essential class of neural-based architectures within our graph learning journey. GNNs are distinctively designed to process graph-structured data, enabling them to leverage the inherent relationships and interconnections embedded within such datasets. Their ability to operate on graph-structured data is different from that of conventional neural networks—which are not inherently designed to interact directly with data that lacks a regular, grid-like structure. Unlike standardized data formats, such as images or sequential text that fit neatly into these traditional models, graph data is inherently complex and irregular, consisting of nodes (or vertices) and edges that represent a diverse range of relationships and attributes.

GNNs have the ability to model and interpret these complex relationships in graph data such as social networks, knowledge graphs, and recommendation systems. Through a process of iterative message passing—in which nodes exchange information and encode these interactions through multiple layers in a manner similar to traditional deep learning models—GNNs are capable of learning detailed and new representations of not just individual nodes and edges but also for the entire graph. The representations derived from GNNs can be used for a wide range of downstream graph learning tasks such as node classification, where each node receives a specific label; link prediction, ...