Graph Neural Network

• Graph neural networks are neural network architectures specifically designed for learning representations of graph-structured data including learning node representation.

  More formally the problem solved by GNNs can be framed as follows. Let be a graph and be the Adjacency Matrix. We also let be the attribute matrix on the node features.

  The goal is to learn a node representation in dimensions, denoted such that the graph structural information and node attributes are preserved.

Supervised Learning §

Misc §

• The following challenges are apparent in graph representation learning and motivate advances in GNNs.
  • High Computational complexity which limits its applicability to large networks.
  • Low Parallelizability since edges imply coupling between nodes.
  • Non-stationarity - samples from graph data are dependent on each other.

Links §

• Graph Neural Networks — Foundations, Frontiers and Applications by Wu et al.

• Graph Theory