GatedGraphConv implementation on link prediction

[Haoyuan-L]

Hi, I tried to implement the above architecture from the paper [https://arxiv.org/abs/1511.05493](Gated Graph Sequence Neural Networks). But I'm confused with the PyG GatedGraphConv operator, is this operator works as a single GG-NN in the network?(e.g. like Fo&Fx in the above figure)

I didn't see implementations of the node annotation output model for predicting X(k+1) from H(k,T) in the source code. This is an essential step for the next node prediction. If I set the out_channels equal to the input size, how can I identify the next link node?

Input and output for the GatedGraphConv operator. (each node has 3 features)

[rusty1s]

Yes, our GatedGraphConv model only implements the "Propagation Model" as given in Section 3.2. You can use these layers to construct the GGS-NN model. Let me know if that resolves your concerns.

[Haoyuan-L]

@rusty1s Thank you for your reply. I still don't understand the output of the operator GatedGraphConv. Is this a hidden state for the next node or node score? If it is a hidden state for the next step, how can it be used for node selection (i.e., identify the node)?

In the above figure, how can the output of GatedGraphConv feed into the output model like ov = g(hv(T) , xv) ?

[rusty1s]

GatedGraphConv will just return new node embeddings, e.g., it starts from an embedding h_v^(1) and produces h_v^(T). For the node selection, take a look at the aggr.AttentionalAggregation operator.

[Haoyuan-L]

Hi @rusty1s , I'm kind of confused with this part of the source code for `GatedGraphConv`. Why do you use a for loop here but only output once? I noticed that the iteration times of the for-loop equal to the sequence length, does the purpose of the loop here is to do the message propagation based on the prediction length?

[rusty1s]

There is no sequence prediction integrated here, this is just the the propagation model described in Section 3. For sequence prediction, you need to follow Section 4, e.g., a GatedGraphConv propagation model is used to predict the next token in the sequence. Hope that clarifies your concerns.

[Haoyuan-L]

Thank you for your reply. Should I remove the for-loop in the forward function to use GatedGraphConv for the next token prediction? (In the path-finding problem, because there is only one timestep between each connected node in the graph, right?)

[rusty1s]

I think you would need to keep it. The proposed model is designed to perform multiple steps of propagation.

[Haoyuan-L]

Thank you, one more question, should I add an aggr.AttentionalAggregation layer after GatedGraphConv to do the sequence prediction?

[rusty1s]

Yes, I think so.