Track: Track1; Team name: weeyev; Model: HyMN

[weeyev]

Description

This PR implements the paper Balancing Efficiency and Expressiveness: Subgraph GNNs with Walk-Based Centrality by Southern et al. for the TDL 2026 Challenge. HyMN uses walk-based centrality scores for both the subgraph selection policy and the structural encoding inputs, increasing the expressiveness of the subgraph GNN without sacrificing efficiency by considering only the top-ranked nodes.

In this implementation, truncated centrality structural encodings are computed from closed walks of lengths 1–16. The two highest-centrality nodes define marked graph views, which are processed alongside the original unmarked graph by shared residual GIN layers. The resulting node representations are aggregated across views and passed to TopoBench’s task-specific readout, supporting both community detection and triangle counting.

Tests

• Added focused unit tests for centrality encodings, marked-node selection, view aggregation, batching, caching, permutation equivariance, gradients, and configuration validation.
• Added graph/hymn to the required smoke-test coverage.
• Verified linting, formatting, focused unit tests, and pipeline smoke-test execution.

Checklist

• Clear and explanatory pull request title
• Ruff lint and formatting checks pass
• Focused unit tests and the pipeline smoke test pass
• PEP8-compliant implementation with NumPy-style documentation

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB