> ML_LITERATURE // HAMILTON-2017-INDUCTIVE-REPRESENTATION-LEARNING-LARGE-GRAPHS_v1.0
Inductive Representation Learning on Large Graphs (GraphSAGE)
William L. Hamilton, Rex Ying, Jure Leskovec · Advances in Neural Information Processing Systems (NeurIPS) (2017)
seminal-architecture2017industry-standardthirdPartyReproduced
Principal Contribution
Introduced fixed-size uniform neighborhood sampling and general aggregation functions (Mean, LSTM, Pooling) for inductive node embeddings.
Operational Relevance
The backbone for industrial graph neural network deployments on billion-scale graphs at Pinterest (PinSage), Amazon, and Uber.
Assumptions
- Uniformly sampling a fixed budget of neighbors retains sufficient structural signal while bounding mini-batch computation
Limitations
- Multi-hop neighbor sampling suffers from exponential neighbor explosion (S_1 * S_2 * ... * S_L) across deep layers