---

name: node-property-prediction description: "Routes OGB node-property workflows for ogbn datasets and node-level evaluators." metadata: disco-role: operating disable-model-invocation: true license: MIT

Node Property Prediction

Use this subskill for the node-property family: ogbn-* datasets, including heterogeneous graphs and the one-graph node-classification loaders.

Use this subskill when

• The task names NodePropPredDataset, PygNodePropPredDataset, DglNodePropPredDataset, or ogbn-*.
• The task asks for node-level ROC-AUC or accuracy evaluation.
• The task mentions ogbn-mag, ogbn-products, ogbn-arxiv, ogbn-proteins, or ogbn-papers100M.

First decisions

1. Read references/workflows.md for the family workflow and hetero-graph caveats.
2. Read references/api-reference.md for the public class names and split/label shapes.
3. Read references/troubleshooting.md when a split, label, or import fails.
4. If the task also asks for packaging a new dataset, jump to dataset-contribution instead of stretching this subskill.

Main workflow

• Pick the exact ogbn-* dataset name.
• Load the dataset and inspect dataset[0] plus get_idx_split().
• Decide whether the task is homogeneous or heterogeneous.
• Use the evaluator's expected input format before building metrics.
• Treat ogbn-mag as a heterogeneous one-graph workflow.
• Treat ogbn-papers100M as the large binary/raw-format workflow.

Common routing choices

• ogbn-arxiv -> citation graph node classification.
• ogbn-products -> large-scale product graph classification.
• ogbn-proteins -> binary node classification with ROC-AUC.
• ogbn-mag -> heterogeneous graph workflow with dict-valued labels/splits.
• ogbn-papers100M -> large binary dataset workflow.

Optional backend note

PyG and DGL wrappers are optional. If they are not installed, the library- agnostic loader remains the right path for the core OGB node workflow.

What not to do here

• Do not route graph-property or link-property tasks here just because they also use graphs.
• Do not assume every ogbn-* dataset returns the same label shape.
• Do not depend on the original checkout for final runtime guidance.

Related references

• ../../references/api-overview.md
• ../../references/dataset-catalog.md
• ../../references/troubleshooting.md