Graph neural networks

How this question breaks down

1. 1Graph basicsCore16 minTurn "who follows whom and who saved what" into a graph, then into numbers a model can use (shared-neighbour counts, counts that discount indiscriminate neighbours, random-walk scores and learned node vectors), and see where each one stops working.
2. 2Message passingCore17 minA GNN layer lets every node collect its neighbours' vectors, combine them with a function that ignores order, and update its own vector. Stack K layers and each node sees K hops, until too many layers blur everyone together.
3. 3At scaleCore20 minFollow one training step and one nightly job through a 34-million-node graph: a step for 1,024 makers reads 87 MB with a [15, 10] fan-out instead of 10.9 GB, and a nightly run that goes layer by layer does about 8.5× less work than one that embeds makers one at a time. A nearline path covers the hours in between.

Work through the parts in order and you’ve covered the full design. Each part is one topic.