HeteroGNNExplainer with HeteroGraphConv

Questions
#1

I have a RGCN model implemented follows

Basically, it uses HeteroGraphConv class.

How can I use this HeteroGNNExplainer
https://docs.dgl.ai/en/1.0.x/generated/dgl.nn.pytorch.explain.HeteroGNNExplainer.html
to obtain the edge masks? Thanks.

#2

You need to modify the model class to incorporate the use of edge weights, as suggested in the example of HeteroGNNExplainer.

#3

I am a bit confused by the weight and edge_weight in graphconv class, dgl.nn.pytorch.conv.graphconv — DGL 1.0.2 documentation

So below is how the RGCN is defined,

 self.conv = dglnn.HeteroGraphConv(
            {
                rel: dglnn.GraphConv(
                    in_feat, out_feat, norm="right", weight=False, bias=False
                )
                for rel in rel_names
            }
        )

 self.weight = nn.Parameter(
                    th.Tensor(len(self.rel_names), in_feat, out_feat)
                )
                nn.init.xavier_uniform_(
                    self.weight, gain=nn.init.calculate_gain("relu")
                )

weight = self.basis() if self.use_basis else self.weight
            wdict = {
                self.rel_names[i]: {"weight": w.squeeze(0)}
                for i, w in enumerate(th.split(weight, 1, dim=0))
            }

hs = self.conv(g, inputs, mod_kwargs=wdict)

So this weight here is basically Wr in RGCN formula, and it is passed to graphconv, where it has two parameters weight=None, edge_weight=None.
I am confused which parameter actually takes this wdict in your example. It seems it is weight, if that is the case, can I simply change wdict to

           wdict = {
                self.rel_names[i]: {"edge_weight": mask}
                for i, w in enumerate(masks)
            }

and then pass this to the trained model?

#4

The weight argument is for employing an alternative weight matrix W in message passing AXW. In the example for HeteroGNNExplainer, it is eweight for edge weights or equivalently the nonzero entries in A. Your modification looks fine to me.

1 Like
#5

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.