icdm_graph_competition/openhgnn_example

• OpenHGNN example

OpenHGNN example

• Dataset process
  Since openhgnn is based on DGL, so the dataset processing of openhgnn is identical to DGL. Refer to DGL EXAMPLE

• Install openhgnn:

git clone -b icdm https://github.com/BUPT-GAMMA/OpenHGNN
cd OpenHGNN
pip install .

• train on session1 dataset.

python session1.py

• inference on session2 dataset with loaded model (trained from session1).

python session2.py

• Parameter Descriptions:

  • Parameters in ICDM2022Dataset:
    • session: The dataset name in the competition, options: "session1", "session2".
    • force_reload: Whether to reload the dataset. Default: False
    • load_labels: Whether to load labels to the graph. For "session2", there is no labels. Default: True
  • Parameters in Experiment:
    • model: The model you use in OpenHGNN, default: "RGCN".
    • dataset: The dataset of "AsNodeClassificationDataset".
    • task: "node_classification" do not change.
    • gpu: The ID of GPU you use. If you want to use cpu, please set to -1.
    • mini_batch_flag: We suggest to set it "True", as the dataset is large.
    • max_epoch: The epoch you use when training.
    • load_from_pretrained: If you already have a .pt file, you can put it under directory ./openhgnn/output/{model} .
    • data_cpu: Whether data is loaded to the CPU to save GPU resources.
    • batch_size: Used when "mini_batch_flag" is "True".
    • direct_inference: Skip the training process and make inferences directly. Please set "load_from_pretrained" to "True".

• Change "session" to change the dataset.

• If your "session" is "session2", please make sure that "load_labels" is "False", "direct_inference" is "True".