Track: Track2; Team name: Remy; Model: WHNN

[stefano-blando]

TDL Challenge 2026 Submission

• Track: Track 2 (TNN)
• Team Name: Remy
• Model Name: WHNN (Wasserstein Hypergraph Neural Network)
• Domain: hypergraph
• Paper: Wasserstein Hypergraph Neural Network (Iulia Duta, Pietro Liò, 2025, arXiv:2506.09682)

Summary of Implementation

• Backbone: Implemented in topobench/nn/backbones/hypergraph/whnn.py. Features 100% vectorized PyTorch native SlicedWassersteinPooling without Python loop overhead for maximum GPU throughput.
• Config: Added Hydra model config in configs/model/hypergraph/whnn.yaml.
• Unit Tests: Full unit test coverage added in test/nn/backbones/hypergraph/test_whnn.py.
• Official Evaluation: Completed the full GraphUniverse benchmark (72 runs, seeds 42/43/44) via official notebook 2026_tdl_challenge/run_evaluation.ipynb.
• Results: Produced and included 2026_tdl_challenge/outputs/2026-08-01_14-37-42/results.json along with heatmaps and OOD plots.

Verification Checklist

• Correctness of model implementation against paper specification
• Fully vectorized PyTorch operations
• All unit tests pass
• Pipeline tests pass
• results.json generated for all 72 GraphUniverse runs