some problem on the accuracy

[Tadelaide]

some problem on the accuracy

[VHRanger]

Hi,

What method from OpenNe are you using? My node2vec implementation generally gets similar link prediction and graph clustering results to the neighbourhood based node embedding implementations from karateclub (another github graph embedding library).

The big difference might be the default parameters you're using. I default to shorter walks with more epochs (faster since generating walks is parallelized) but others often have fewer, longer walks, which may be better for some graph structures.

If you are planning to embed huge graphs, I would be patient, there's a release in the works which lets you avoid using networkX altogether (reading an edgelist directly to a compact CSRGraph and embedding from it).

I've had a few questions about embedding huge graphs, and generally networkX won't scale to huge graphs because you'll run out of RAM past a few hundred thousand nodes.

[VHRanger]

Closing since I can't replicate the issue. See the link prediction notebook in the examples folder for reproduction code.

I'll reopen happily if we can replicate the performance issue.