FastRP embeddings

Wiki / Graph data science

These examples require the people projection from the projection guide.

FastRP embeddings

FastRP initialises sparse random vectors, optionally appends projected numeric features, propagates weighted neighbourhood averages, and sums L2-normalised intermediate vectors. embeddingDimension is required and positive. iterationWeights defaults to [0.0,1.0,1.0]; its length controls the number of propagation steps. nodeSelfInfluence defaults to zero and weights the initial normalised vector. Empty iteration weights require nonzero self influence. normalizationStrength (default zero) scales initial vectors by outgoing degree raised to that power; isolated nodes use degree one for this scaling.

propertyRatio is in [0,1]; the property portion has floor(embeddingDimension*propertyRatio) coordinates. A positive ratio requires nonempty featureProperties, whose values must be finite numeric scalars or nonempty numeric lists with consistent dimensions. Feature vectors are shared across nodes. At ratio one with a fixed seed, initialisation depends on features and not numeric node IDs. Remaining coordinates use node ID and seed. A fixed seed is reproducible within GDB; RNG sequences are implementation-specific. An isolate has a zero embedding unless self influence contributes. The final weighted sum is not normalised again. Overflow raises an error.

CALL gds.fastRP.mutate('people', {
  embeddingDimension: 128, randomSeed: 42, mutateProperty: 'structuralEmbedding'
});
CALL gds.knn.stream('people', {
  nodeProperties: 'structuralEmbedding', topK: 10, randomSeed: 42
}) YIELD node1, node2, similarity;
CALL gds.graph.nodeProperties.write('people', ['structuralEmbedding']);

Related articles

Graph projections and catalogue · GDS algorithms and execution modes · GDS resources and cancellation · Graph data science