Category: B2; Team name: DLLB; Dataset: PPI

[dleko11]

Co-authored-by: luka-benic luka121benic@gmail.com
Co-authored-by: dleko11 davidleko11@gmail.com

Checklist

• My pull request has a clear and explanatory title.
• My pull request passes the Linting test.
• I added appropriate unit tests and I made sure the code passes all unit tests. (refer to comment below)
• My PR follows PEP8 guidelines. (refer to comment below)
• My code is properly documented, using numpy docs conventions, and I made sure the documentation renders properly.
• I linked to issues and PRs that are relevant to this PR.

Description

This PR extends TopoBench to support edge-level link prediction on both transductive and inductive graph datasets, and adds a tutorial notebook that illustrates how to use the new functionality.

Concretely, the PR introduces:

• Edge-level split utilities for link prediction (transductive and inductive).
• A dynamic negative sampling transform integrated into the dataloading pipeline.
• A dedicated edge-level readout (LinkPredictionReadOut) for link prediction on top of existing GNN backbones (GCN, GAT).
• Example dataset configurations for Cora (transductive), MUTAG (inductive), and PPI (inductive, predefined splits).
• A tutorial notebook showing the full workflow end-to-end.

---

Key Changes (Code)

• Edge-level splitting

  • load_edge_transductive_splits for single-graph / transductive datasets (e.g. Cora).
  • load_edge_inductive_splits for multi-graph / inductive datasets (e.g. MUTAG, PPI).
  • Both return DataloadDataset objects with:
    • edge_label_index, edge_label (positive and negative candidate edges),
    • consistent handling of val/test negatives vs train-time negatives.

• Dynamic negative sampling

  • NegativeSamplingTransform in topobench.transforms.data_manipulations:
    • takes positive edges from edge_label_index,
    • samples fresh negatives via torch_geometric.utils.negative_sampling,
    • rebuilds edge_label_index / edge_label each epoch according to neg_pos_ratio and neg_sampling_method.

• Edge-level readout

  • LinkPredictionReadOut in topobench.nn.readouts:
    • consumes node embeddings x_0 from the backbone,
    • scores candidate edges via dot products,
    • outputs 2-class logits (no-edge, edge) and attaches labels for the loss/evaluator.

• PPI dataset support

  • New loader (based on torch_geometric.datasets.PPI) that:
    • loads the predefined train/val/test splits from PyG,
    • combines them into a single dataset with a split_idx mapping,
    • is compatible with the inductive edge-level splitting utilities.

• Configuration-level support

  • task_level: edge and num_classes: 2 for link prediction.
  • Extended split_params for link prediction:
    • learning_setting (transductive / inductive),
    • val_prop, test_prop, train_prop,
    • is_undirected,
    • neg_pos_ratio (dynamic train negatives),
    • neg_sampling_ratio (static val/test negatives),
    • neg_sampling_method.

These changes plug into the existing TopoBench training pipeline without altering the high-level interface (Hydra configs + run.yaml).

---

Tutorial (Usage Example)

To illustrate the new link prediction support, this PR also adds:

• tutorials/tutorial_link_prediction.ipynb

The notebook demonstrates:

• Transductive and inductive link prediction setups demonstrated on the Cora and MUTAG datasets.
• How the split utilities, negative sampling transform, and LinkPredictionReadOut interact in practice.
• Running short GCN/GAT experiments and inspecting basic metrics and visualizations of positive/negative edges in the splits.

The tutorial is an example user guide for the new functionality; all core logic lives in the library code.

Issue

Additional context

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[luka-benic]

We fixed some compatibility issues, namely we had a problem with the PPI dataset class from torch_geometric version 2.8.0. which was not compatible with networkx version 2.8.8.