Modify just some nodes embeddings based in index

Questions
#1

Hi, everyone.

I’m following the graph sage approach for link prediction: 5.3 Link Prediction — DGL 0.6.1 documentation

My graph is homogeneous, one link type, one node type.

My question is: could I just sage some part of the nodes during training? I have 80000 nodes, but I would need to modify node embeddings from index 40000 to 80000.

Is this possible?

Thanks.

#2

It suffices to remove all the incoming edges of nodes with index 0 to 39999, so their embeddings will never be updated by message passing. You also need to properly set self-loops or residual connections for them.

#3

Since all my origin nodes are from 0 index to 40000, and all my destination nodes are from 40000 to 80000, I dont need to remove any edge, right?

Sorry, but I dont fully understand this part.

Thank you so much.

#4

Since all my origin nodes are from 0 index to 40000, and all my destination nodes are from 40000 to 80000, I dont need to remove any edge, right?

Yes

Sorry, but I dont fully understand this part.

With AX for message passing, where A is the adjacency matrix and X is the node feature matrix, nodes without incoming edges will have a full-zero feature. You need to carefully handle that.

#5

All my nodes from 40000 to 80000 have al leats one incoming edge, so this doesn´t affect them, right? Or maybe it does, since the origin nodes from those edges didn’t have themselves upcoming edges?

#6

Sorry for the late reply. If your GNN has more than one GNN layers, then you do need to have self-loops or residual connections for nodes from 0 to 39999.

1 Like
#7

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