A graph view of interactions
A recommender's data is naturally a bipartite graph: users on one side, items on the other, with an edge for each interaction. Graph based methods exploit this structure so signal flows along connections, reaching items several hops away.
Message passing
A graph neural network updates each node by aggregating its neighbors.
- A user vector gathers from the items it touched.
- An item vector gathers from the users who touched it.
- Stacking layers lets a user reach items liked by similar users two hops out.
This multi hop spread captures collaborative signal that direct neighbors alone miss.
A light recipe
Some popular models drop the heavy nonlinear transforms and just propagate and average embeddings across the graph, then sum the layer outputs. Simpler often matches or beats complex variants here.
Strengths and costs
- It surfaces relevant but distant items through chains of similarity.
- Large graphs need neighbor sampling to keep training tractable.
Key idea
Graph based recommenders model interactions as a bipartite graph and propagate embeddings across edges, letting multi hop message passing surface collaborative signal beyond direct neighbors.