About dataset

[graph-0]

About dataset

[GRAPH-0]

Hi @Syzseisus .
Thanks for your interest!
We extract node structural features using the node2vec and node semantic features using Transformer-XL.
Refer to Paper:

[Syzseisus]

Thank you for your quick response.

I've already reviewed the provided resources.
However, I'm interested in obtaining a detailed understanding of the configuration settings for
node2vec and Transformer-XL models.
Could you please provide me with the specific parameters or something?

ALTERNATIVELY, for FB15k, I would appreciate it if you could confirm whether the order of nodes in the pickle file provided in Google Drive is the same as in the dgl.
If you are unsure, please let me know the order
or you can check it by following these steps:

1. For the order of nodes used in dgl, you can access the raw tgz file from here
   and find the order in the entities.dict file.
2. However, please note that 0dgl uses the node entity name as a hash.
   To find the real names of the nodes, you can refer to the entities2wikidata.json file.

Thank you for your assistance.

Sincerely,
Wooseong Cho.

[GRAPH-0]

Hi,

1. For node embedding, you can refer to #3 and #6.

The node2vec settings:

• FB and tmdb:

model = Node2Vec(data.edge_index, embedding_dim=64, walk_length=20,
                     context_size=10, walks_per_node=10, num_negative_samples=1,
                     p=1, q=1, sparse=True).cuda()
loader = model.loader(batch_size=128, shuffle=True, num_workers=4)
optimizer = torch.optim.SparseAdam(list(model.parameters()), lr=0.01)

• imdb:

model = Node2Vec(data.edge_index, embedding_dim=128, walk_length=20,
                     context_size=10, walks_per_node=10, num_negative_samples=1,
                     p=1, q=1, sparse=True).cuda()

2. Sorry, I fail to open the tgz file from dgl to check the order. But you can get the correspondence between node id and mid from datasets/FB15k/fb15k_description.tsv. I think this can help you!

Sincerely,
Han