GraphSAGE

• http://snap.stanford.edu/graphsage/

Machine learning with graphs

Implementations

• Original: https://github.com/williamleif/GraphSAGE

It allows “identity features” option to work on static graphs without features.

If your benchmark/task does not require generalizing to unseen data, we recommend you try setting the “–identity_dim” flag to a value in the range $$64,256$$ . This flag will make the model embed unique node ids as attributes, which will increase the runtime and number of parameters but also potentially increase the performance. Note that you should set this flag and not try to pass dense one-hot vectors as features (due to sparsity). The “dimension” of identity features specifies how many parameters there are per node in the sparse identity-feature lookup table.

Explanations

• Graph SAGE - Inductive Representation Learning on Large Graphs | GNN Paper Explained