Message Passing Neural Networks Framework
What you'll learn: How the Message Passing Neural Networks (MPNN) framework provides a unified mathematical structure for understanding all GNN variants.
The Unifying View
You've now seen GCN, GraphSAGE, and GAT — different architectures with different strengths. But they all share a common pattern. The MPNN framework formalizes this shared structure, showing that every graph neural network can be described using three core functions:
- Message function φ: How to compute messages from neighbors
- Aggregation function ⊕: How to combine messages (sum, mean, max, attention-weighted)
- Update function γ: How to compute the new node representation
The General Formula
For each node v at layer k, the MPNN framework defines:
m_v^(k) = ⊕_{u ∈ N(v)} φ(h_v^(k-1), h_u^(k-1), e_{u,v})
h_v^(k) = γ(h_v^(k-1), m_v^(k))
Where:
- φ creates messages from neighbor features and optional edge features
- ⊕ aggregates all neighbor messages (permutation invariant)
- γ updates the node embedding using old state and aggregated messages
How Your Models Fit
- GCN: φ is identity with normalization, ⊕ is normalized sum, γ applies weights and activation
- GraphSAGE: φ is identity, ⊕ can be mean/max/LSTM, γ concatenates and transforms
- GAT: φ computes attention scores, ⊕ is attention-weighted sum, γ applies final transformation
Think of MPNN as the "recipe template" — every GNN is just filling in different ingredients for these three functions.
Key Takeaway: The MPNN framework unifies all GNN architectures as variations of three composable functions: message, aggregation, and update — making it easier to understand, compare, and design new graph neural networks.