Which loss could deal with unsupervised link prediction of heterogeneous graph?

[wanglu2014]

The main question is the loss function. Thank you for your help!

[rusty1s]

I guess the loss formulation for link prediction in heterogeneous graphs is the same for link prediction in homogeneous graphs: You would need to utilize a contrastive loss formulation, e.g., via the negative log likelihood of positive and negative sampled examples.

[wanglu2014]

Thank you for your timely reply. Our goal is to focus on edges between two types of nodes. Did you have implementation?

[rusty1s]

Don't have an implementation for that, sorry, but I guess you can come up with something quite quickly by looking at examples/link_pred.py and a heterogeneous graph model, e.g., this one.