Supervised GraphSAGE and a faster Unsupervised GraphSAGE in Aura Graph Analytics

You can often tell more about a node from its connections than from the node itself: a fraudulent account rarely transacts alone. GraphSAGE is a graph neural network that learns from those connections as well as each node’s own properties. In Aura Graph Analytics it now comes in two dedicated flavors: one for classifying nodes and one for embedding them.

Supervised GraphSAGE uses the labels you already have to make predictions for the rest of your graph. You train a model on a set of labeled nodes, such as accounts already confirmed as fraudulent. It then learns to classify every other node from its properties and its neighborhood. Built-in validation and class weighting help the model hold up even when the class you care about is rare. You can stream, mutate or write the predictions with (Python) and (Cypher).

Unsupervised GraphSAGE turns every node into an embedding: a compact vector that captures what the node is and how it’s connected. Embeddings are ready to use for similarity search, clustering, or your own ML models. No labels are needed, and training is much faster than with the existing GraphSAGE endpoint. It's available through its own endpoint, (Python) and (Cypher).

Both endpoints are available in Preview in Aura Graph Analytics as of the 2026.10 release, via Cypher and Python (with GDS Python client 2.2 or later). Your existing GraphSAGE workflows are unaffected.

Supervised docs · Unsupervised docs