OpenHGNN/openhgnn/tasks

• Task
• • • • Included Object:
      • Method:
      • Supported task
      • • Node classification
        • Link prediction
    • How to build a new trainerflow

Task

The task means downstream task. So it is related to dataset closely.

Included Object:

• Dataset, containing the heterogeneous graph as DGLGraph, as well as node/edge features and additional dataset-specific information.
• Evaluation Metric, a callable object which takes in the prediction and the ground truth label, and returns an output.
  • Currently we will implement evaluation metrics for node classification and link prediction.
• Training Loss, a callable object which takes in the prediction and the ground truth as well as any other necessary information and returns an output. It must be differentiable.
  • Ranking loss for self-supervision, or contrastive learning losses, all fall into this category.
• Training-validation-test Split.

Method:

• get_graph()
  • return the DGLGraph
• get_loss_fn()
  • train with a full_batch graph
• evaluate()
  • Evaluate Metric
• get_labels()
  • return labels

Supported task

• Node classification

• Link prediction

• Recommendation

How to build a new trainerflow

1. Create a class your_task that inherits the BaseTask and register the task with @register_task(str).
2. We decorate the func evaluate() with @abstractmethod. So the evaluate() must be overridden, or the your_task cannot be instantiated.
3. Fill the dict SUPPORTED_TASKS in task/init.py