Playground experiment

GNN Message Passing Toy

Choose node features, apply one to three message-passing layers, and watch neighboring embeddings become smoother without leaving the browser.

← Back to Playground Graph Signal Diffusion

Interactive GNN

Average neighbor messages

Start from editable node features, then repeatedly replace each embedding with the average of itself and its neighbors. More layers mix farther neighborhoods and reduce edge-level contrast.

Edge variation ready

Feature preset

Node feature Node 1

Message-passing layers

Before

edge variation 0.00

After

edge variation 0.00

Embedding plot

2 layers

What changes per layer

Each layer expands the receptive field by one hop. A node first blends direct neighbors, then neighbors of neighbors, then a wider local patch of the graph.

This is useful because nearby context can improve a representation, but too many layers can make connected nodes look too similar.

What the colors mean

The initial colors are hand-picked toy features A, B, and C. The after view shows the learned embeddings after aggregation, so mixed colors mean a node has absorbed neighboring feature channels.

The embedding plot uses an equilateral A/B/C feature triangle. Pure features sit on the corners, while mixed embeddings move into the interior.