Create Heterogeneous Graphs

[DuanhaoranCC]

Dear author, Hello! I'm building my own heterogeneous graph dataset, but I seem to have a problem. I have two kinds of nodes A (200), B (450), and 75 different types of edges connecting A and B.
A_list = [0,1,...199] B_list = [0,1,...449] data['A', 'relation1', 'B'].edge_index = torch.tensor(np.array([A_list[5], B_list[365], [A_list[56], B_list[2]]), dtype=torch.long).t().contiguous() data['A', 'relation2', 'B'].edge_index = torch.tensor(np.array([A_list[15], B_list[35], [A_list[9], B_list[78]]), dtype=torch.long).t().contiguous()
But in doing so, I am reporting an error that the dimension exceeds the index in the input model. I would like to ask if this composition is correct?

[rusty1s]

I assume you create the graph as follows?

import numpy as np
import torch

from torch_geometric.data import HeteroData

A_list = list(range(200))
B_list = list(range(500))

data = HeteroData()
data['A', 'relation1', 'B'].edge_index = torch.tensor(
    np.array([[A_list[5], B_list[365]], [A_list[56], B_list[2]]]),
    dtype=torch.long).t().contiguous()
data['A', 'relation2', 'B'].edge_index = torch.tensor(
    np.array([[A_list[15], B_list[35]], [A_list[9], B_list[78]]]),
    dtype=torch.long).t().contiguous()

print(data)
print(data['A', 'relation1', 'B'].edge_index)
print(data['A', 'relation2', 'B'].edge_index)

This should work just fine. Can you show me where your script crashes?

[DuanhaoranCC]

We want to get the meta-path of the following figure according to the official documentation. Is there any way? Because the heterogeneous graph attention needs to be calculated through the meta-path, and I only have:

data['paper', 'cites', 'paper'].edge_index = ... # [2, num_edges_cites]
data['author', 'writes', 'paper'].edge_index = ... # [2, num_edges_writes]
data['author', 'affiliated_with', 'institution'].edge_index = ... # [2, num_edges_affiliated]
data['paper', 'has_topic', 'field_of_study'].edge_index=...

This doesn't appear to be a meta path. Please what should I do to get the meta path like DBLP:

data[('p', 'a')].edge_index = nei_p
data[('a', 'p', 'a')].edge_index = apa
data[('a', 'pcp', 'a')].edge_index = apcpa
data[('a', 'ptp', 'a')].edge_index = aptpa
data[('a', 'pos', 'a')].edge_index = pos

Thanks for your guidance!

[rusty1s]

You can add arbitrary meta-paths to your HeteroData object via the AddMetaPaths transform. Let me know if this resolves your concerns :)

[DuanhaoranCC]

Thank you for your suggestion! I am building the above image (https://github.com/DuanhaoranCC/CUB/blob/main/KG.png).
Our node type and edge type are:

from torch_geometric.transforms import AddMetaPaths

data = torch.load('data.pt')
metapaths = [[("A", "B"), ("B", "A")]]
data = AddMetaPaths(metapaths)(data)
['A', 'B']
[('A', 'chin', 'B'), ('B', 'chin', 'A'), 
('A', 'base', 'B'), ('B', 'base', 'A'), 
('A', 'remex', 'B'), ('B', 'remex', 'A')]

DBLP:

['author', 'paper', 'term', 'conference']
[('author', 'to', 'paper'), ('paper', 'to', 'author'), 
('paper', 'to', 'term'), ('paper', 'to', 'conference'), 
('term', 'to', 'paper'), ('conference', 'to', 'paper')]

We compared the difference between the two, We find

"C:\Users\Lenovo\AppData\Local\Programs\Python\Python37\lib\site-packages\torch_geometric\transforms\add_metapaths.py", line 107

for j, metapath in enumerate(metapaths):
    for edge_type in metapath:
        print(data._to_canonical(edge_type))
        assert data._to_canonical(
            edge_type) in data.edge_types, f"'{edge_type}' not present"

it will turn ("A","B") side into ('A', 'to', 'B'), but ('A', 'to', 'B') not in data.edge_types, so an error occurred:

Traceback (most recent call last):
  File "F:/GRAND/2.py", line 97, in <module>
    data = AddMetaPaths(metapaths)(data)
  File "C:\Users\Lenovo\AppData\Local\Programs\Python\Python37\lib\site-packages\torch_geometric\transforms\add_metapaths.py", line 107, in __call__
    edge_type) in edge_types, f"'{edge_type}' not present"
AssertionError: '('A', 'B')' not present

[rusty1s]

Since there is no unique relation between A and B, you will need to define the relation type when defining edge types, e.g.:

metapaths = [[("A", "chin", "B"), ("B", "chin", "A")]]

[DuanhaoranCC]

Thank you again for your patient guidance. All the problems have been solved, which makes me have an in-depth understanding of Pytorch Geometry.