gds.fastRP.mutate creates all zeros!?

Neo4j Graph Platform
,
1

Hi, I created a graph with Document nodes and Word(lemmatized) nodes which is contained in the document, so Documents are connected through common Word nodes.

I projected the part of them like this:
call gds.graph.project.cypher(
'documentRootword',
'
match (n) where (n:Document and date("20220416")<=n.datePublished<=date("20220515")) or n:RootWord
return id(n) as id
',
'
match (d)-[:FROM_DOCUMENT]->(r:RootWord) where date("20220416")<=d.datePublished<=date("20220515")
return id(d) as source, id(r) as target
'
)

Then run fastRP like this:
call gds.fastRP.write(
'documentRootword',
{
embeddingDimension: 128,
iterationWeights: [1.0, 0.5, 0.5],
normalizationStrength: 0,
writeProperty: 'embedRootword',
randomSeed: 7
}
)

I know fastRP is designed for homogeneous graph. But I can usually create embeddings in heterogeous graph like this before and get useful similarities in them when I did it with GDS last year.

match (d:Document) where date("20220416")<=d.datePublished<=date("20220515")
return d.embedRootword limit 100
d.embedRootword
[0.049813542515039444, -0.033209025859832764, -0.06641805171966553, 0.03320902958512306, 0.016604511067271233, 3.2323665966060844e-9, 0.09962708503007889, -0.08302256464958191, 0.03320902958512306, 3.2323665966060844e-9, 0.14944063127040863, 3.2323665966060844e-9, 0.033209025859832764, 0.033209025859832764, 0.016604512929916382, -0.14944063127040863, -0.09962708503007889, 0.049813542515039444, -0.049813542515039444, 0.04981353506445885, 0.049813542515039444, -0.08302256464958191, 0.049813542515039444, -0.03320903703570366, -0.11623159050941467, 0.06641805171966553, 0.049813542515039444, 0.09962708503007889, 0.049813542515039444, 3.2323665966060844e-9, 0.11623159050941467, 0.23246318101882935, 0.13283610343933105, 0.06641805171966553, 0.06641805171966553, -0.033209025859832764, -0.016604512929916382, 0.0, -0.08302255719900131, -0.01660451665520668, -0.03320902958512306, -0.19925417006015778, 0.01660451665520668, 0.14944063127040863, -0.09962708503007889, -0.03320902958512306, -0.016604511067271233, -0.16604512929916382, 0.03320903703570366, 0.049813542515039444, -0.03320902958512306, 0.03320902958512306, -0.04981353506445885, 0.0830225721001625, -0.24906770884990692, -0.049813542515039444, 0.016604511067271233, 0.1826496422290802, 0.24906770884990692, 0.0830225721001625, -0.033209025859832764, 0.016604511067271233, -0.016604511067271233, 0.21585866808891296, -0.09962708503007889, -0.049813542515039444, 9.697100011862858e-9, -0.01660451665520668, -0.03320902958512306, -0.08302256464958191, 0.03320902958512306, -0.049813542515039444, 0.03320902958512306, -0.11623158305883408, -0.06641805171966553, 0.11623159050941467, 0.06641805171966553, -0.16604512929916382, 0.09962707757949829, -0.24906770884990692, -0.11623158305883408, -0.11623159050941467, -0.049813542515039444, -0.1826496422290802, 0.049813542515039444, 0.06641805171966553, 0.01660451665520668, 0.11623158305883408, -0.016604511067271233, -0.01660451665520668, 0.03320902958512306, 0.08302255719900131, 0.0, -0.049813542515039444, 0.09962708503007889, -0.049813542515039444, 0.049813542515039444, 0.033209025859832764, 0.016604511067271233, -0.03320903703570366, -3.2323665966060844e-9, 0.016604511067271233, -0.049813542515039444, 0.016604512929916382, -0.03320902958512306, 0.049813542515039444, 0.033209025859832764, -0.09962707757949829, 0.13283610343933105, 0.06641805917024612, 0.0830225721001625, 0.08302256464958191, -0.08302256464958191, -0.03320902958512306, 3.2323665966060844e-9, 0.06641805171966553, -0.14944063127040863, -0.06641805917024612, 0.0, -0.03320903703570366, 0.16604512929916382, -0.11623159050941467, -0.06641805171966553, -0.016604511067271233, -3.2323665966060844e-9, -0.033209025859832764, -0.0830225721001625, -0.08302256464958191]
....

But now with GDS 2.0.3 or 2.0.4, I cannot get the same embeddings in the memory from mutate procedure even though I can get the embeddins by Write procedure:
call gds.fastRP.mutate(
'documentRootword',
{
embeddingDimension: 128,
iterationWeights: [1.0, 0.5, 0.5],
normalizationStrength: 0,
mutateProperty: 'embedRootword',
randomSeed: 7
}
)

call gds.graph.streamNodeProperty(
'documentRootword',
'embedRootword'
)
yield nodeId, propertyValue
return nodeId, propertyValue limit 10
==> It returns all zeros for the values in the embeddings.

nodeId

propertyValue

25

[0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]

27

[0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]

...

Am I doing something wrong?

2

hi @gigauser ,

This problem has been bugging me for the past few days and I think I have found the solution for my case.

I have changed the relationship orientations to 'UNDIRECTED' when projecting a graph. This somehow fixed the issue. I guess FastRP prefers 'UNDIRECTED' graphs as it is mentioned in the docs.

3

I landed here as I too was getting mostly vectors of all zeros (though some had other floats)

For me, setting the NodeSelfInfluence did the trick.

My objective is just to preserve monopartite hop distances as euclidean distances in the embedding.