Graph Neural Networks, or GNNs, learn from data that is connected as a graph. Instead of only rows, images, or sequences, the data contains nodes and edges: people connected to friends, products connected to buyers, molecules connected by bonds, or accounts connected by transactions.
Why it matters
Many high-value problems are relational. Fraud rings, recommendation networks, supply chains, knowledge graphs, proteins, and social platforms depend on connections. GNNs help models learn from both the entity and its neighborhood.
Key terms
- Node: an entity such as a user, product, molecule atom, or web page.
- Edge: a relationship between nodes, such as follows, bought, transferred-to, or bonded-with.
- Node feature: attributes attached to a node.
- Adjacency: the connection structure of the graph.
- Message passing: nodes send information to neighbors and update their representations.
- Node embedding: learned vector that represents a node and its graph context.
- Link prediction: predicting whether an edge should exist between two nodes.
- Over-smoothing: deep GNN layers can make node embeddings too similar.
How message passing works
- Start with node features.
- Each node collects messages from its neighbors.
- Messages are aggregated using mean, sum, attention, or another function.
- The node updates its embedding using its old state and the aggregated neighborhood signal.
- After several rounds, each node contains information from a wider graph neighborhood.
Visual explanation suggestion
Show nodes as glowing points connected by edges. Let learners select a node, then animate one-hop and two-hop message passing with color intensity showing how neighbor information changes the selected node embedding.
Common mistakes
- Using a GNN when the relationships are weak, noisy, or irrelevant.
- Ignoring edge direction, edge type, or edge weight when those details matter.
- Building too many layers and causing over-smoothing.
- Creating train/test leakage by splitting nodes without respecting time or connected components.
- Forgetting that graph construction choices can dominate model quality.
Interview-style questions
- What kinds of problems are a good fit for graph neural networks?
- Explain message passing in simple terms.
- What is the difference between node classification and link prediction?
- Why can train/test splitting be tricky for graph data?
Related lessons
- Vectors & Scalars
- Neural Network Architecture
- Attention Mechanism
- Model Evaluation Metrics
- AI Ethics & Bias
Related project/template CTA
A graph-based fraud or recommendation starter can extend the MLOps Starter Kit when you are ready to build a production portfolio project.