Add TemporalGraphConv

[aurorarossi]

This PR adds a TemporalGraphConv constructor to each snapshot of the TemporalSnapshotsGNNGraph, a convolutional homogeneous graph layer.
A PR will follow to update the documentation.

[CarloLucibello]

Paper citation is missing, and also more explanation of how it works.

Am I correct in saying that this applies layer independently to each temporal slice?

[aurorarossi]

Yes, you are right.
This constructor has been used in the following papers, for example

• https://dl.acm.org/doi/pdf/10.1145/3583780.3615091

• https://www.semanticscholar.org/reader/e9b4aeb1371256aa5cf907621935b40266318e78

• https://proceedings.neurips.cc/paper_files/paper/2021/file/22785dd2577be2ce28ef79febe80db10-Paper.pdf

And I do not know of any paper that defines it. I would say it is similar to the HeteroGraphConv for heterographs.

[CarloLucibello]

I would say it is similar to the HeteroGraphConv for heterographs.

Yes, although HeteroGraphConv in the end aggregates messages coming from different edge types, while here there is no cross-talk between different times, it has to happen in separate modules (e.g. temporal attention in the last reference above)

[CarloLucibello]

Suggested change

	return tgconv.layer.(tsg.snapshots, x)
	return tgconv.layer.(tsg.snapshots, x)

Also, here broadcasting should be avoided, since both snapshots and x could be expanded

[aurorarossi]

Suggested change

	return tgconv.layer.(tsg.snapshots, x)
	return map((a, b) -> tgconv.layer(a, b), tsg.snapshots, x)

Is this a better way to write the broadcast in this case?

[CarloLucibello]

sorry, I got confused, it is the same provided x is a vector of the same length as snapshots.

[CarloLucibello]

I think that instead of introducing a new layer, we can extend all layers to take a temporalsnapshotgraph as input and apply itself timewise. The time dimension in this way is like a batch dimension.
It would be something like

function (l::GNNLayer)(g::TemporalSnapshotsGNNGraph, x::AbstractVector; kws...)
    y = ....
    return y
end  

Maybe the operation could be made more efficient batching the snapshots first and then unbatching the output y later, but this is a performance optimization that can be explored later.

[aurorarossi]

Ah ok I see. Do you prefer one big PR with support for TemporalSnapshotsGNNGraphs for all layers or one PR for each layer?
In both cases I will close this PR.
Thanks!