Alternate weight matrix in a bipartite graph

Questions
#1

Hi,

I’m doing some experimentations in my research in which I will need to use different weight matrixes for different node types in a heterogeneous bipartite graph.

For the sake of example, in message massing and aggregation, every node of type A must use a weight matrix W, and every node of type B must use a weight matrix W’. Nodes type A doesn’t have any interference in the weight matrix W’ and vice versa.

Is it possible to implement this using DGL?

#2

Below is an example and see if it meets your need.

import dgl
import torch
from dgl.nn.pytorch import HeteroGraphConv, GraphConv

# Create a bipartite graph of two node types A and B
g = dgl.heterograph({
      ('A', 'A->B', 'B'): (torch.tensor([1, 2]), torch.tensor([2, 3])),
      ('B', 'B->A', 'A'): (torch.tensor([2, 2]), torch.tensor([1, 3]))
})
# Set initial node features
g.ndata['h'] = {'A': torch.randn(4, 1), 'B': torch.randn(4, 1)}
# Use a GraphConv for each relation. In a bipartite graph, this is 
# equivalent to a weight matrix per node type
model = HeteroGraphConv({'A->B': GraphConv(1, 2, norm='none', bias=False),
                         'B->A': GraphConv(1, 2, norm='none', bias=False)})
model(g, g.ndata['h'])
1 Like