music-graph-clustering.cypher

// 1. Add placeholder properties onto non-song entities (needed for FastRP)

MATCH (n)
WHERE n:Artist OR n:Genre OR n:Album
SET n.instrumentalness = 0.0,
    n.speechiness = 0.0,
    n.tempo = 0.0,
    n.danceability = 0.0,
    n.acousticness = 0.0,
    n.instrumentalness = 0.0,
    n.liveness = 0.0,
    n.valence = 0.0,
    n.loudness = 0.0,
    n.energy = 0.0,
    n.popularity = 0.0;

// 2. Project graph into session

WITH gds.aura.api.credentials("aTCwhbP2B7BkZMenb1K2Opunlx0vQYnk", "...") AS credentials
MATCH (source:Song)-[r]->(target)
WITH gds.graph.project('musicgraph', source, target, {
  sourceNodeLabels: "Song",
  targetNodeLabels: labels(target),
  sourceNodeProperties: source {
     .instrumentalness, .speechiness, .tempo, .danceability, .acousticness,
    .liveness, .durationMs, .valence, .loudness,
    .energy, .popularity
  },
  targetNodeProperties: target {
    .instrumentalness, .speechiness, .tempo, .danceability, .acousticness,
    .liveness, .valence, .loudness,
    .energy, .popularity
  },
  relationshipProperties: { }
}, {
  memory: '4GB',
  ttl: duration({minutes: 5})
}) AS g
RETURN g.graphName, g.nodeCount, g.relationshipCount;

// 3. Run FastRP

WITH gds.aura.api.credentials("aTCwhbP2B7BkZMenb1K2Opunlx0vQYnk", "...") AS credentials
CALL gds.fastRP.mutate('musicgraph', {
  randomSeed: 42,
  embeddingDimension: 8,
  mutateProperty: 'embedding',
  iterationWeights: [0.8, 1, 1, 1],
  normalizationStrength: 0.5,
  featureProperties: [
    'instrumentalness','tempo', 'danceability', 'acousticness',
    'liveness', 'valence', 'loudness',
    'energy', 'popularity'
  ]
})
YIELD nodePropertiesWritten
RETURN nodePropertiesWritten;

// 4. Filter the graph to only Song nodes with embedding

WITH gds.aura.api.credentials("aTCwhbP2B7BkZMenb1K2Opunlx0vQYnk", "...") AS credentials
CALL gds.graph.filter(
  'musicgraph2',        // New graph name
  'musicgraph',         // Existing in-memory graph
  'n:Song',             // Node filter: only Song nodes
   '*'
)
YIELD graphName, nodeCount, relationshipCount
RETURN graphName, nodeCount, relationshipCount;

// 5. Run KMeans to cluster songs based on embedding.

WITH gds.aura.api.credentials("aTCwhbP2B7BkZMenb1K2Opunlx0vQYnk", "...") AS credentials
CALL gds.kmeans.write('musicgraph2', {
  nodeProperty: 'embedding',
  writeProperty: 'cluster',
  k: 40
}) YIELD computeMillis
RETURN computeMillis;

// 6. Extract cluster as separate node (playlists)

MATCH (s:Song)
WHERE s.cluster IS NOT NULL
MERGE (c:Playlist {id: s.cluster})
MERGE (s)-[:IN_PLAYLIST]->(c);

// 7. Clean up temporary properties.

MATCH (n)
WHERE n:Artist OR n:Genre OR n:Album
REMOVE n.instrumentalness,
    n.speechiness,
    n.tempo,
    n.danceability,
    n.acousticness,
    n.liveness,
    n.durationMs,
    n.key,
    n.timeSignature,
    n.valence,
    n.loudness,
    n.energy,
    n.embedding,
    n.popularity;

// 8. Drop projection

WITH gds.aura.api.credentials("aTCwhbP2B7BkZMenb1K2Opunlx0vQYnk", "...") AS credentials
CALL gds.graph.drop("musicgraph")
YIELD schema
RETURN *;

WITH gds.aura.api.credentials("aTCwhbP2B7BkZMenb1K2Opunlx0vQYnk", "...") AS credentials
CALL gds.graph.drop("musicgraph2")
YIELD schema
RETURN *;