Seq-HGNN: Learning Sequential Node Representation on Heterogeneous Graph

Requirements

System: Linux x86_64, Nvidia GPU (>=8GB GPU memory).

We recommend employing Conda to install the runtime environment.

• cuda==11.* (In my case, cuda==11.7)

• python==3.10

  conda create -n seq_hgnn python=3.10
  conda activate seq_hgnn

• pytorch==2.0.1

  conda install pytorch=2.0.1 torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia

• pyg==2.3.0

  conda install pyg=2.3.0 -c pyg
  conda install pytorch-sparse -c pyg

• dgl==0.9.1

  conda install -c dglteam dgl-cuda11.7=0.9.1post1

• dglke==0.1.0 (install from source)

  pip install git+https://github.com/awslabs/dgl-ke.git@43875a57a721d62396d05235226b13e6c2db1a2a#subdirectory=python

• ogb

  conda install -c conda-forge ogb

• einops

  conda install -c conda-forge einops

• torchmetrics==0.10

  conda install -c conda-forge torchmetrics=0.10

• wandb

  conda install -c conda-forge wandb

Data preparation

ogbn-mag

In this part, we refer to NARS and SeHGNN.

First, we generate node embeddings using dgl-ke.

cd data/complex_nars
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

Training

ogbn-mag

• Vanilla Seq-HGNN (without Label Propagation & Multi-Stage Training)

  Under this root folder, run

  python -m seq_hgnn.train --lr 0.0005 --num-hidden 512 --num-layers 2 --batch-size 256 --n-batch 250 --dropout 0.5 --num-samples 1800 --dataset ogbn-mag-complex --logsubfix vanilla --epochs 200 --workers 8 --device 0 --amp --off-wandb

• Seq-HGNN + Label Propagation

  python -m seq_hgnn.train_lp --lr 0.0005 --num-hidden 512 --num-layers 2 --batch-size 256 --n-batch 250 --dropout 0.5 --num-samples 1800 --dataset ogbn-mag-complex --logsubfix lp --epochs 200 --workers 8 --device 0 --amp --off-wandb

• Seq-HGNN + Label Propagation + Multi-Stage Training

  python -m seq_hgnn.train_lp_ms --lr 0.0005 --num-hidden 512 --num-layers 2 --batch-size 256 --n-batch 250 --dropout 0.5 --num-samples 1800 --dataset ogbn-mag-complex --amp --logsubfix ms --save-model --stage 200 75 75 75 75 --ct 0.55 --device 0 --off-wandb