How to peform inductive inference using Shadow-GNN about shadow_gnn HOT 1 OPEN

For this you need to prepare separate adj files, adj_train that defines the connectivity among training nodes only and adj_full that defines connectivity among nodes seen or unseen during training. adj_full will be used during the validation & test phases.

For example, on the Flickr graph, we perform inductive training (I think it aligns with your use case). In the config yaml you can specify that the training is inductive:

https://github.com/facebookresearch/shaDow_GNN/blob/main/config_train/flickr/vanilla/sage_3_ppr.yml#L3

Then when loading the graph for training, the model will separately load adj_train and adj_full, and information of adj_full is unseen during training:

https://github.com/facebookresearch/shaDow_GNN/blob/main/config_train/flickr/vanilla/sage_3_ppr.yml#L3

Here is an example of the data format conversion script (GraphSAINT format --> shaDow format): https://github.com/facebookresearch/shaDow_GNN/blob/main/para_graph_sampler/graph_engine/frontend/loader.py#L43

In this section, we defined all the data files required to train shaDow-GNN (you don't need to prepare those optional files, as they will be auto-generated):

https://github.com/facebookresearch/shaDow_GNN#general-shadow-format

Please let me know of any additional questions. Thanks.

from shadow_gnn.