Skip to content

Inductive Representation Learning on Large Graphs (GraphSAGE)

Authors: William L. Hamilton, Rex Ying, Jure Leskovec Year: 2017 Citations: 15,000+

Summary

GraphSAGE (Sample and Aggregate) enables inductive GNN learning on unseen nodes. Samples and aggregates from neighborhoods instead of computing all node embeddings.

Key Concepts

  • Inductive Learning: Learn embeddings for unseen nodes
  • Neighborhood Sampling: Sample and aggregate K-hop neighbors
  • Aggregation Functions: Mean, LSTM, pooling combine features
  • Mini-Batch Training: Efficient sampling-based training
  • Scalability: Works on massive graphs beyond memory
  • Generalization: Learned aggregations transfer to new nodes

Impact

  • 15,000+ citations
  • Enabled inductive learning on graphs
  • More practical than transductive GCN
  • Better scalability for production systems
  • Influenced subsequent GNN designs
  • Foundation for production GNNs
  • Applications in recommendation, social networks
  • Graph Convolutional Networks (Kipf & Welling, 2017)
  • Graph Attention Networks (Veličković et al., 2017)
  • Node2Vec (Grover & Leskovec, 2016)