use unsupervised GraphSAGE(SAGEConv) to classify nodes when only edges have features?

[kokifish]

Is there any way to use unsupervised GraphSAGE to classify nodes when only edges have features?
In more detail:

• nodes: without features
• edges: directed with features
• goal: node classification
  When node A sampling A's neighbour node B, B's features are the features of directed edge A->B.
  When node C sampling C's neighbour node B, B's features are the features of directed edge C->B.
  Node A/C can sample node B only if directed edge A/C->B exits.

In other words, I have a Data like:
Data(x=[num_nodes, target_embed_dim], edge_index=[2, num_edges], edge_attr=[num_edges, num_edge_features])
where x is a tensor of all zeros.
Can I do this by reimplementing SAGEConv? If yes, how to reimplement it?

[rusty1s]

There are two options to tackle this:

1. You need to utilize a GNN tack can incorporate edge features, e.g., TransformerConv (see here).
2. You need to first convert your edge features into node-level ones, e.g. via scatter_mean(edge_attr, edge_index[1], ...). This can potentially also be learned as part of a GNN pipeline (scatter_mean(self.mlp(edge_attr), ...))

[kokifish]

Thank you for your patience.

If I understand correctly. Filtering of edge features means filter_data, more precisely, filter_edge_store_. Thus the Data return by NeighborLoader only holds the edge features for neighbours that are present in the sampled graph(Data).

My doubt is whether SAGEConv's forward(propagate() in forward) can only aggregate the node features(Data.x as used here) of neighbouring nodes, and cannot use edge features(Data.edge_attr) as node features.

Do I have to reimplement the forward propagate() for passing edge_attr and message_and_aggregate() for using edge_attr for matmul in SAGEConv?
Very much looking forward to your reply.

[rusty1s]

Yes, that is correct. SAGEConv cannot handle edge features by definition. As an alternative, you can use any GNN layer that can handle edge features, see here. If you want to incorporate edge features into SAGEConv, this will look as follows (note that message_and_aggregate cannot be implemented in this scenario):

def forward(self, x, edge_index, edge_attr):
    out = self.propagate(edge_index, x=x, edge_attr=edge_attr):
    return self.lin2(x) + out

def message(self, x_j, edge_attr):
    return self.lin1(torch.cat([x_j, edge_attr], dim=-1))

[kokifish]

Thank you very much for your patience.
It works for me. Thanks again!

[downeykking]

Yes, that is correct. SAGEConv cannot handle edge features by definition. As an alternative, you can use any GNN layer that can handle edge features, see here. If you want to incorporate edge features into SAGEConv, this will look as follows (note that message_and_aggregate cannot be implemented in this scenario):

def forward(self, x, edge_index, edge_attr):
    out = self.propagate(edge_index, x=x, edge_attr=edge_attr):
    return self.lin2(x) + out

def message(self, x_j, edge_attr):
    return self.lin1(torch.cat([x_j, edge_attr], dim=-1))

Hello, I am confused by this

def forward(self, x, edge_index, edge_attr):
    out = self.propagate(edge_index, x=x, edge_attr=edge_attr):
    return self.lin2(x) + out

def message(self, x_j, edge_attr):
    return self.lin1(torch.cat([x_j, edge_attr], dim=-1))

when we add edge_attr in sageconv, should this be changed to

def forward(self, x, edge_index, edge_attr):
    out = self.propagate(edge_index, x=x, edge_attr=edge_attr):
    return self.lin2(x) + out

def message(self, x_j, edge_attr_j):
    return self.lin1(torch.cat([x_j, edge_attr_j], dim=-1))

because the x_j has been chosen by x[row], so maybe edge_attr should be transfered to edge_attr_j?

[downeykking]

Yes, that is correct. SAGEConv cannot handle edge features by definition. As an alternative, you can use any GNN layer that can handle edge features, see here. If you want to incorporate edge features into SAGEConv, this will look as follows (note that message_and_aggregate cannot be implemented in this scenario):

def forward(self, x, edge_index, edge_attr):
    out = self.propagate(edge_index, x=x, edge_attr=edge_attr):
    return self.lin2(x) + out

def message(self, x_j, edge_attr):
    return self.lin1(torch.cat([x_j, edge_attr], dim=-1))

Hello, I am confused by this

def forward(self, x, edge_index, edge_attr):
    out = self.propagate(edge_index, x=x, edge_attr=edge_attr):
    return self.lin2(x) + out

def message(self, x_j, edge_attr):
    return self.lin1(torch.cat([x_j, edge_attr], dim=-1))

when we add edge_attr in sageconv, should this be changed to

def forward(self, x, edge_index, edge_attr):
    out = self.propagate(edge_index, x=x, edge_attr=edge_attr):
    return self.lin2(x) + out

def message(self, x_j, edge_attr_j):
    return self.lin1(torch.cat([x_j, edge_attr_j], dim=-1))

because the x_j has been chosen by x[row], so maybe edge_attr should be transfered to edge_attr_j?

I got it. This is right.

def forward(self, x, edge_index, edge_attr):
    out = self.propagate(edge_index, x=x, edge_attr=edge_attr):
    return self.lin2(x) + out

def message(self, x_j, edge_attr):
    return self.lin1(torch.cat([x_j, edge_attr], dim=-1))

Thanks!