Huon Wilson

I've made a start at #1860, but there's still some issues to work through that I haven't been able to do today.

@dokato noted in #1920 that the import can crash with "illegal hardware instruction" on M1.

A link prediction task is often used to predict link existence. Many of the other examples in https://stellargraph.readthedocs.io/en/stable/demos/link-prediction/index.html do this. Instead of adding all of the didn't watch edges, the...

Hi, thanks for filing an issue! GAT is quite similar to GCN, so I'd recommend starting with [the GCN link prediction demo](https://stellargraph.readthedocs.io/en/stable/demos/link-prediction/gcn-link-prediction.html) and changing the `GCN(...)` model to `GAT(...)`. You...

Ah, I think this is a bug in TensorFlow: https://github.com/tensorflow/tensorflow/issues/41361. Changing `metrics=["acc"]` to `metrics=["binary_accuracy"]` results in non-zero accuracies.

Hi, thanks for asking a question. It sounds like there's two things here: - The GML format might be most easily loaded via NetworkX: https://stellargraph.readthedocs.io/en/stable/demos/basics/loading-networkx.html - Loading data for graph...

Unfortunately this seems to duplicate a bunch of the node-data/edge-data interaction work from #1220, unlucky! 🙁 A 29× speed-up is pretty nice! > Things do get a bit slower for...

> That's an interesting idea. So we would build the index automatically when graph.neighbors is called with a non-None other_node_type value? Yeah. > As an alternative, maybe it could be...

A 3× speedup still seems worth it, but yeah, maybe we can leave it till after the 1.0 crunch.

Hm, 2 hours with this PR or with `develop`? #1446 is related too.