Node embeddings
Node embedding algorithms compute low-dimensional vector representations of nodes in a graph. These vectors, also called embeddings, can be used for machine learning and vector search. The Neo4j Graph Data Science library contains the following node embedding algorithms:
Using node embeddings
Node embeddings support two common workflows:
-
Machine learning: Use embeddings as input features for downstream tasks such as node classification, link prediction, and kNN similarity graph construction.
-
Structural similarity search: Write the embeddings to node properties, create a vector index in the Neo4j database over that property, and retrieve nodes that are similar by graph topology. This structural vector search can be useful on its own or combined with lexical and semantic results in hybrid search.
See the Developer Guides for more examples of hybrid search.
Generalization across graphs
In machine learning tasks, often the graph used for constructing the embeddings and training the downstream model differs from the graph on which predictions are made. Compared to normal machine learning where we just have a stream of independent examples from some distribution, we now have graphs that are used to generate a set of labeled examples. Therefore, we must ensure that the set of training examples is representative of the set of labeled examples derived from the prediction graph. For this to work, certain things are required of the embedding algorithm, and we denote such algorithms as inductive [1].
In the GDS library the algorithms
are inductive.
Embedding algorithms that are not inductive we call transductive. Their usage should be limited to the case where the test graph and predict graph are the same. An example of such an algorithm is Node2Vec.