Redundant self connections in ogbn-products

[tedzhouhk]

Dear OGB Team,

Thanks for providing these awesome datasets!

I noticed that for the ogbn-products dataset, there are some redundant self connections. For example, the edge (15064,15064) occurred twice in the edge_index array.

Here is a list of the row and column indices of the repeated edges. After removing the repeated edges, it gives me 123718152 (doubled) edges instead of the original 123718280 edges.

Best,
Tedzhouhk

[weihua916]

Thank you very much for pointing this out. Indeed, a very small number of self-edges do exist in the raw data (which we did not expect), and the issue here is that we have added inverse edges for all the edges (including the self-edges) when making the graph undirected. Given that self-edges are very rare (128 nodes out of 2,449,029 nodes), we are not planning to release a modified dataset, but it is good to keep this in mind. Please feel free to remove the duplicated self-edges.

[weihua916]

We have added a brief note about this issue on our webpage. Thanks again for noticing this!

[tedzhouhk]

Thank you so much for your quick response!