Graph Embeddings

ML models eat tensors, not adjacency lists. Graph embeddings map each node (and sometimes each relation) to a vector so that position encodes structure — nodes with similar neighborhoods land close together.

That unlocks four capabilities traversal can't give you: link prediction ('which edges should exist but don't'), node classification, similarity over structure rather than text, and candidate generation for entity resolution. The tradeoff is inherent: a vector is a lossy summary of a neighborhood — you get a score, not the path that produced it.