When visualizing graph nodes, should I use apply PCA to node2vec embedding?

Question

I am trying to visualize graph nodes using node2vec embedding. The node2vec embeddings has lengths of 50~100 dimensions.

I have two plans:

1. use umap to project node2vec embeddings to 2D space
2. use PCA to project node2vec embeddings to a slightly lower-D space (~30-50 dimensions), and then use umap for 2D space outputs.

Which plan is better?

Answer 1

The PCA step will remove redundancies: this is a solid first step. Then as you say it can be followed up with the 2D embedding. Option 2.

Answer 2

Do the Correlation analysis first IMHO if you find highly related features you may drop some, once done you may perform PCA on the remaining features, this will help you reduce the feature while keeping the maximum variance.