Graph Machine Learning
But vector embeddings of nodes and edges are still an intermediate step. To capture topology as vectors — to encode n-hop context, not just text — you need graph embeddings and graph machine learning. A graph embedding knows about your content and about the structure it sits in.
The field is harder to find than it should be, partly because it goes by too many names (geometric deep learning, network embeddings, graph representation learning, non-Euclidean ML, ML on graphs…). Same effort, different boxes. Over the years I’ve worked with Pykeen, PyG, Spektral, TorchDrug, AutoGL, DeepSNAP, DGL and others — for drug repurposing, clinical trial optimization, even astrophysics.
But not yet in mainstream agentic systems. That’s the gap I find most interesting right now. Agentic stacks are converging on knowledge graphs as memory and reasoning substrate, but they’re still treating those graphs as retrieval surfaces, not as objects to learn on. The bridge between agentic KGs and graph ML isn’t there yet — and yes, large-scale graph ML remains a GPU-intensive investment, which is part of why.
If you’re thinking about what an agent that actually learns on its knowledge graph looks like, I’d like to hear it. I think there’s a serious opportunity hiding in there.