search

dmlc / dgl

Python package built to ease deep learning on graph, on top of existing DL frameworks.
http://dgl.ai
Apache License 2.0
13.54k stars 3.01k forks source link

Why does examples/pytorch/mvgrl/graph/dataset.py compute "Personalized Page Ranking" of graphs? #5679

Closed yurivict closed 1 year ago

yurivict commented 1 year ago

This example computes something called Personalized Page Ranking in the procedure compute_ppr.

What is Personalized Page Ranking of a graph and why is it needed in order to train a neural network on graphs?

I think this testcase needs a README explaining why this is done.

@hengruizhang98 Maybe you know what does compute_ppr compute and why is it needed to train a neural network on graphs?

hengruizhang98 commented 1 year ago

This is the paper's design. See Eq.3 in the paper.

yurivict commented 1 year ago

I see , thank you.

Also the example prints accuracy_mean, 0.8985 before training, and accuracy_mean, 0.8193 after training. Does this mean that training didn't succeed, since accuracy didn't improve much?

Also the loss function is negative, which is strange.

Also this example doesn't have any test set to evaluate accuracy.

hengruizhang98 commented 1 year ago

This can happen in self-supervised node/graph representation learning tasks, as there are some studies claiming that an untrained GNN can already perform quite well. If the self-supervised loss function is not consistent with the downstream tasks, pertaining can harm the performance.

There is train/test split in: https://github.com/dmlc/dgl/blob/master/examples/pytorch/mvgrl/graph/utils.py#L12-L26

hengruizhang98 commented 1 year ago

The loss can be negative as it is a difference between positive loss and negative loss. https://github.com/dmlc/dgl/blob/master/examples/pytorch/mvgrl/graph/utils.py#L63-L83

If the positives' scores are lower than negatives' scores, it will become negative.

github-actions[bot] commented 1 year ago

This issue has been automatically marked as stale due to lack of activity. It will be closed if no further activity occurs. Thank you

frozenbugs commented 1 year ago

Hi @yurivict , I am closing this issue assuming you are happy about our response. Feel free to follow up and reopen the issue if you have more questions with regard to our response.