Link Prediction Pipeline Experiment[Help please] - Node Embeddings Comparision

Neo4j Graph Platform Graph Algorithms/Graph Data Science
1

Hello all,

I am conducting an experiment to compare the different node embeddings in Link Prediction Pipelines and I ran into some errors - hopefully someone can point out what I did wrong.

Ran FastRP and Node2vec - no problems

image
image940×379 102 KB

image
image940×373 86.8 KB

However, ran into this problem with GraphSage and HashGNN

image
image940×396 126 KB

GraphSage:

image
image1391×476 48.3 KB

so created separate model...but

image
image1376×591 54.2 KB

Trying to make sense of the documentation but :sob: feeling so dumb at the moment

2

Hi @rosinialexander,

For HashGNN: HashGNN - Neo4j Graph Data Science
You will need to specify the two required parameters iterations and embeddingDensity according to the doc. In addition, since there is no featureProperties (if it is specified, that feature will be used as the initialisation of hashGNN embeddings), you need to also specified a generateFeatures mapping that tells the algorithms how to initialise embeddings.

For GraphSAGE:
As you pointed out, GraphSAGE needs to be trained separately and not as a step in the pipline. GraphSAGE - Neo4j Graph Data Science
The error of your separate model suggests that the features (Novel, Session) are not on the nodes in the projected graph Dev1. I think you'll need to reproject a the graph with these features on them. Then GraphSAGE knows it should initialise all node embeddings as those features.

Thanks,
Brian

3

Hi @brian.shi1

Thank you so much for your advice - will reproject the Graph and try again.