Node embedding training on large graph

[garyhsu29]

Hello, I have a large dataset that contains around 200,000,000 nodes and 3,000,000,000 edges. In my case, it's a single large graph and the graph is dynamic. My goal is to learn node embedding for downstream classification task or use some clustering algorithms to find the community of nodes.

I have two questions here:

1. Because the graph is so large and I could not load the whole graph into my RAM, and I am wondering how to split my whole graph into multiple subgraphs without losing a lot of information about its topology.
2. What would be the suitable algorithm that I should use in my case? Do you think the graphsage + edge reconstruction loss could learn the embedding properly?

[rusty1s]

1. Wow, this is a giant graph. We are currently working on letting PyG scale to out-of-memory datasets via FeatureStore and GraphStore abstractions (will be part of the next release). Currently, you would need to rely on external sources to split your graph before inputting into PyG.

2. Yes, that sounds good. GraphSAGE is always a strong baseline.

[garyhsu29]

Thank you so much!