Node Similarity and KNN

Use an existing projection; KNN property names must have been projected or added by mutate.

Node Similarity compares outgoing neighbour sets using JACCARD, OVERLAP or COSINE. KNN compares property values: scalar numbers use 1/(1+abs(a-b)); integer lists default to Jaccard; floating lists default to normalised cosine. Integer lists also support OVERLAP, and floating lists support EUCLIDEAN and normalised PEARSON. Multiple property scores are averaged. Property configuration accepts strings and metric maps:

CALL gds.knn.stream('people', {
  nodeProperties: [{embedding: 'COSINE'}], topK: 10, similarityCutoff: 0.8
}) YIELD node1, node2, similarity;

KNN prepares numeric vectors and integer sets once. It uses sampled neighbour descent: start with uniformly sampled neighbours, refine using forward and bounded reverse neighbour lists, and add random candidates each iteration. Defaults are sampleRate:0.5, maxIterations:100, deltaThreshold:0.001, randomJoins:10. The first two fractions must be in (0,1]; iterations must be positive and random joins non-negative. Set randomSeed for reproducibility. nodeProperties also accepts a single property name. Similarity cutoff is applied after neighbour refinement so it cannot block candidate exploration.

The GDB extension exact:true performs exhaustive comparisons. Graphs with at most 64 selected nodes, or K covering every other node, automatically use exact search. Stats report exact, nodePairsConsidered, ranIterations and didConverge; convergence means neighbour updates met the stopping threshold, not guaranteed exact neighbours. Approximate results may change from the older exact-only implementation; use exact:true when exact results are required. Ties prefer the lower node ID. RNG sequences are implementation-specific; random-walk initial sampling and a perturbation option are not supported.

Node Similarity remains exact but retains at most K neighbours per source while comparing pairs, giving O(NK) retained candidates instead of O(N^2). KNN also retains O(NK) neighbour state, plus prepared property data and a per-node candidate set. For positive similarity cutoffs and more than 64 selected nodes, Node Similarity can enumerate only pairs sharing a neighbour. A cheap join-work estimate selects this path on sparse graphs; dense/hub-heavy graphs and zero-cutoff queries retain exhaustive comparison. Exact scores, zero-score behaviour and node-ID tie ordering are preserved. Worst-case comparison work remains quadratic.

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