A Graph Neural Network (GNN) is a neural network designed to process data represented as a graph. Graphs are useful when relationships between entities are as important as the entities themselves.
A graph consists of:
1. Nodes: Represent entities such as people, products, molecules, or locations.
2. Edges: Represent relationships or connections between nodes.
3. Node features: Describe individual nodes.
4. Edge features: Describe relationships, when such information is available.
GNNs commonly use a message-passing process. During each layer, a node receives information from its neighboring nodes, combines those messages using an aggregation function, and updates its own representation. After several layers, a node can contain information from nodes that are multiple edges away.
A general message-passing layer can be described in three stages:
1. Message generation: Neighboring nodes produce messages based on their current representations and possibly edge features.
2. Aggregation: Messages are combined using an operation such as summation, averaging, or maximum pooling.
3. Update: The node combines the aggregated information with its existing representation to produce a new representation.
Different GNN architectures use different message and aggregation functions. Graph Convolutional Networks (GCNs), Graph Attention Networks (GATs), and GraphSAGE are important examples.
GNNs can support node-level, edge-level, and graph-level prediction tasks. For example, a model may classify individual nodes, predict whether a relationship should exist, or classify an entire molecular graph. Challenges include scalability on very large graphs, over-smoothing across layers, and limited information propagation over long graph distances.
Graph Neural Networks
Graph Neural Networks are neural architectures designed to learn from graph-structured data by exchanging and combining information between connected nodes.
h_v^(l+1) = UPDATE(h_v^(l), AGGREGATE({h_u^(l) : u ∈ N(v)}))
Explanation
Variables
| Symbol | Meaning |
|---|---|
h_v^(l) | Representation of node v at layer l |
h_v^(l+1) | Updated representation of node v |
N(v) | Set of neighboring nodes connected to node v |
h_u^(l) | Representation of neighboring node u at layer l |
AGGREGATE | Function that combines information from neighboring nodes |
UPDATE | Function that calculates the new node representation |
l | GNN layer index |
Example or application
In a social network, each person can be represented as a node and each friendship as an edge. A GNN can combine information about a person's friends to learn representations useful for predicting interests or recommending content.
Text
GNNs represent data as nodes and edges. Through message passing, each node updates its representation using information from its neighbors, allowing the network to learn relationships within a graph.
Key Points
Graphs represent entities and their relationships.
Nodes represent entities, while edges represent connections.
Message passing allows neighboring nodes to exchange information.
Aggregation combines information from neighboring nodes.
GNNs support node-, edge-, and graph-level prediction.
Different GNN architectures use different message-passing strategies.
Deep GNNs can experience over-smoothing and scalability challenges.
Applications
Social network analysis
Recommendation systems
Molecular property prediction
Drug discovery research
Fraud detection
Traffic network prediction
Knowledge graph reasoning
Learning Outcomes
Explain the structure of graph-structured data.
Identify nodes, edges, and graph features.
Describe the message-passing process.
Differentiate node-, edge-, and graph-level tasks.
Practice Questions
What are nodes and edges in a graph?
How does message passing work in a GNN?
What is the purpose of an aggregation function?
How does a node-level prediction differ from a graph-level prediction?
Related Concepts
Graph Theory
Neural Networks
Message Passing
Graph Convolutional Networks
Graph Attention Networks
Knowledge Graphs
Common Mistakes
Assuming a GNN treats every node independently.
Confusing nodes with edges.
Assuming every GNN uses the same aggregation function.
Ignoring the effect of graph structure on learned representations.
Sources and verification
Sources are shown as provided in the project data. Not every source has yet been independently verified. Please report errors or missing citations.
- Scarselli, F. et al., The Graph Neural Network Model, IEEE Transactions on Neural Networks, 2009.
- Kipf, T. N., and Welling, M., Semi-Supervised Classification with Graph Convolutional Networks, 2017.
- Hamilton, W. L., Graph Representation Learning, Morgan & Claypool, 2020.
Suggest a correction or source · More in Artificial Intelligence & Machine Learning