GraphSage unsupervised loss function

[aure_bnp]

The paper “Inductive Representation Learning on Large Graphs (Graphsage)” has an unsupervised loss function based on random walks (sec 3.2) that is similar to deepwalk and node2vec, I couldn’t find it anywhere.

Is there an implementation or a plan to implement this loss function in dgl?

For instance, for a binary cross entropy loss function you just have to use : torch.nn.functional.binary_cross_entropy.
Is there a similar function for the loss of the authors of the section 3.2 ?

Thank you all !

[minjie]

Is this example helpful?

[aure_bnp]

Not really… There is no implementation of the loss function that I was talking about and of the authors research paper (section 3.2. and see image below).

It seems that I have to code it myself.
But in another post ( GraphSage unsupervised loss function based on randomwalks - Questions - Deep Graph Library (dgl.ai)) you said that an example will be implemented. Did you do it ?

Thanks a lot for your fast reply ! It is not what I wanted but it still can be usefull for implementing the loss function I was talking about.

[aure_bnp]

Hi @minjie, did you find something to help me or does no one have code this loss function yet ?

Thank you.

[mufeili]

I think https://github.com/dmlc/dgl/blob/master/examples/pytorch/graphsage/advanced/train_lightning_unsupervised.py#L66 is very close to the equation you mentioned above, if not exactly the same. You just perform negative log likelihood minimization for positive edges and maximization for negative edges.