🌀🔗 Extend EA datasets to allow loading a unified graph

[mberr]

This PR adds support for loading KGs from entity alignment datasets as a joint graph comprising both individual sides.

[mberr]

@dobraczka - any thoughts on this? maybe also regarding extension to the OpenEA dataset family, or generally adding a workflow to perform EA as LP task?

[dobraczka]

I can take a detailled look at it next week since I have vacation right now. I think treating EA as a special LP task can only get you so far.
Generally, I already did some work in that direction. I used some techniques from the OpenEA package as guidance. The simplest extension is to use "swapping", i.e. from the training alignment add triples to the connected graph, where known matches are exchanges with each other. For example if you have the triples:
A rel1 B
B rel2 C
D rel3 E

and know that A == D you add the triples
D rel1 B
A rel3 E

The other simple strategy is "sharing" where same entites get the same id, but the TriplesFactory does not like that.
For evaluation you then need something like an align-rank-based evaluator, that gets queried with entities from one side and should score the entities from the other side (without the use of a relation). I already hacked something together that does this.

I am currently working on inductive EA. I already created semi- and fully-inductive datasets from the OpenEA datasets and now I'm trying to get a sensible baseline using NodePiece. I got some results already, but it needs a bit of work still.

If you or @migalkin are interested in collaborating on the inductive EA stuff let me know.
As I said I have vacation (until end of next week), so I might be slow to respond in that period, but I am generally very interested in helping to add EA to Pykeen!

[mberr]

Thanks @dobraczka for the quick response despite being on vacation 😅 If you send me your preferred email to max.berrendorf@gmail.com, I'll send you an invite to our PyKEEN Slack.

[dobraczka]

I like that you already included the possibility to extend this with different combination strategies, which is what I was thinking about, but that can be done in a seperate PR.
Do you want to also do this for the OpenEA datasets or should I do that?

[dobraczka]

This train/test/val split would generally only make sense for a LP setting, right? In the EA setting we would want to split the alignment tuples appropriately.
But creating EA-specific settings could also be done in a seperate PR.

[mberr]

I removed some of the additional changes to reduce this PR to a atomic change. The old version is available here.

[mberr]

@PyKEEN-bot test