Bars: AUC of the scores; dots: nodes and edges
Expected common neighbours vs p_out

Link Prediction on a Community Graph: Hidden Edges and Scores (2D)

This simulator builds a random community-structured graph, gives every node a random 3D embedding, and hides a fraction of the real edges as a test set. A GCN-style neighbor-mean aggregation step — the same mechanism behind GCN, GraphSAGE and GAT-style message passing — repeatedly pulls each node's embedding toward the average of its visible (train) neighbors. Score every non-train pair by cosine similarity or raw dot product, sweep a threshold to see which held-out links get correctly predicted versus missed, watch false-positive predictions light up, and track a live AUC estimate as the embeddings converge.

Made with 🟠 Blender