Hi, Folks,
I have been looking into DGL’s graphsage model quality, the pre-built one came in the example folder (dgl/node_classification.py at master · dmlc/dgl · GitHub). I used ogbn-arxiv dataset, when using the default hyper-parameter came with DGL’s implementation, graphsage gets an accuracy of 0.55 on test set. On the contrary, ogb’s leaderboard shows a accuracy of 0.71 for a typical graphsage implementation.
Is this a known issue? I did a bit comparison, but not sure what I missed.
Any suggestions?
Screenshot of ogbn leader board
Output I got from running dgl’s node_classification.py implementation.