SeHGNN on Ogbn-mag

For environment setup, please refer to the main page of this repository.

Training without extra embeddings

python main.py --stages 300 300 300 300 --num-hops 2 --label-feats --num-label-hops 2 --n-layers-1 2 --n-layers-2 2 --residual --act leaky_relu --bns --label-bns --lr 0.001 --weight-decay 0 --threshold 0.75 --patience 100 --gama 10 --amp --seeds 1

For the first time this command is executed, the dataset ogbn-mag will be automatically downloaded in the folder ../data/.

The codes generate random initialized embeddings for node types author/topic/institution.

Training with extra embeddings from ComplEx

For better model effects, we utilize ComplEx to generate embeddings for those three node types with no raw features.

Following NARS, we use dglke to generate embeddings. However, dglke only supports dgl<=0.4.3. We modify some files so that it fits the latest versions of dgl.

1.Generate extra embeddings from ComplEx

Please make sure that the ogbn-mag dataset has been downloaded in the folder ../data/.

Then under the folder ../data/complex_nars, run

python convert_to_triplets.py --dataset mag
bash train_graph_emb.sh mag

Check the running log to find where the generated ComplEx features are saved. For example, if the save folder is ckpts/ComplEx_mag_0, run

python split_node_emb.py --dataset mag --emb-file ckpts/ComplEx_mag_0/mag_ComplEx_entity.npy

2.Training our SeHGNN model

Under this folder, run

python main.py --stages 300 300 300 300 --extra-embedding complex --num-hops 2 --label-feats --num-label-hops 2 --n-layers-1 2 --n-layers-2 2 --residual --act leaky_relu --bns --label-bns --lr 0.001 --weight-decay 0 --threshold 0.75 --patience 100 --gama 10 --amp --seeds 1

Performance