Shouldn't Node2Vec Graph Embedding be done on the original graph instead of clustered graphs?

[ksachdeva]

In phase 3 described here https://microsoft.github.io/graphrag/posts/index/1-default_dataflow/

The node2vec embeddings are shown to be computed after community detection. I have verified that in the code also the implementation follows the documentation

graphrag/graphrag/index/workflows/v1/create_base_entity_graph.py

Line 65 in 9d99f32

"column": "clustered_graph",

This means that node2vec is run on every clustered graph. If there are 3 levels then it is run 3 times.

I do not have a deeper understanding of node2vec (yet!) but I would assume that the embeddings are going to be independent of the community and clustered levels of the graph and rely on the proximity between nodes and edges. Perhaps this is not the case.

Locally I have compared the graph_embedddings of various levels and they all are to be similar/same (cosine distance of 1.0)

Would appreciate some insights.

Regards & thanks
Kapil