Graph Neural Network Message Passing

Node feature update via neighborhood aggregation across L message-passing layers.

Prompt

A graph neural network message-passing diagram.

Top — Input Graph:
- A small graph with 6 nodes and labeled edges. Each node has a feature vector x_i (drawn as a small bar above the node).

Center — Message Passing Layers (stacked, L=3 layers shown):
- For each layer:
  - Each node aggregates features from its neighbors (highlight one focal node and color its neighbors).
  - Aggregation function (mean / sum / max) shown as a small symbol on the aggregation arrow.
  - Update function (small MLP) labeled "phi" at the focal node.
  - The node feature h_i^(l) becomes h_i^(l+1).

Bottom — Readout / Pooling:
- Node features at layer L are pooled (mean or attention-weighted) into a single graph-level embedding.
- The graph embedding feeds an MLP classifier producing the prediction.

Style: clean academic vector, navy / teal palette, sans-serif labels, white background.

Use in Generator

When to use

For GNN papers (GCN, GAT, GraphSAGE, MPNN) and applications like molecular property prediction.

Variations

Graph Attention (GAT)

Replace simple aggregation with attention-weighted aggregation. Show the attention weights alpha_{ij} as varying-thickness edges between focal node and its neighbors.

Tips

Highlight a single focal node per layer — showing all updates at once becomes unreadable.
Annotate aggregation type (mean / sum) explicitly. It changes the model class.
Show the readout step — without it readers don't see how graph-level predictions are made.

FAQ

Can I extend this to a heterogeneous graph?

Add per-edge-type weights and one set of node-update parameters per node type. Color edges by type to make the heterogeneity visible.