Reverse_types for as_edge_prediction_sampler not properly documentated

📚 Documentation

https://docs.dgl.ai/en/0.8.x/generated/dgl.dataloading.as_edge_prediction_sampler.html

This tutorial doesn’t mention that when passing reverse_etypes dict, it is assumed that edges of the reverse types have the same id. (e.g clicks and clicked-by) I found this info here: https://discuss.dgl.ai/t/exclude-eids-in-edgedataloader/1552

[Rhett-Ying]

Let’s add clarity into doc. It’s great if you can help do this.

[as51340]

could you share a minimum and runnable snippet to elaborate this

I don’t have runnable but I can give you an example which I encountered. Let’s say that you are loading a graph from a graph database in which nodes are connected with edges in both directions. The most logical way to convert a graph from graph databases to DGLGraph is to iterate over nodes and their out_edges or in_edges. It is really hard(or more to say computationally inefficient) to search for each edge you encounter, its reverse edge pair and hence construct a graph in which edge and reverse_edge have the same id.