If you would like to evaluate NARS on OAG, please follow instructions below to prepare the dataset. The goal here is to extract graph structure and node features and store into a more efficient format so that we can easily read and convert to a DGLGraph object.

• dill

We use the OAG dataset shared by authors of Heterogeneous Graph Transformer here:
https://github.com/acbull/pyHGT#oag-dataset

Please download the graph_CS.pk from HGT's google drive: https://drive.google.com/drive/folders/1a85skqsMBwnJ151QpurLFSa9o2ymc_rq

ATTENTION! If your python version is >= 3.8, you might get a pickle error since there's some change in serialization. You may need python <= 3.7.

The following steps will take a short while to finish since they all need to load and process a 8.1 GB file.

python3 extract_paper_feats.py

This generates a new file paper.npy.

python3 convert_OAG_venue.py

This generates a new file graph_venue.pk.

python3 convert_OAG_field.py

This generates a new file graph_field.pk.

You might wonder why the graph structures for venue prediction and field prediction are different.
The reason is that in the original OAG graph shared by HGT authors, paper nodes are connected to ground truth venue nodes and field nodes. In order to avoid any chance of information leakage, for venue prediction task, we remove edges between papers and venues, and for L1-field prediction tasks, we remove edges between papers and L1-fields.