# Coauthor Physics

**Task**: Research-field classification (co-authorship)  
**Size band**: medium · **Label type**: semantic  
**Label column**: `field` · **Converter**: `converters/convert_npz.py`

Microsoft Academic co-authorship graph in physics; 5 field classes; the largest shchur benchmark (34.5k nodes).

## Converted graphs (neext/)

| graph | nodes | edges | classes | feature cols | isolated | class counts |
|---|---|---|---|---|---|---|
| default | 34,493 | 247,962 | 5 | 0 | 0 | 2: 17,426, 0: 5,750, 1: 5,045, 4: 3,519, 3: 2,753 |

*Conversion notes*: shchur gnn-benchmark npz; CSR adjacency symmetrized. Feature matrix too large for CSV (skipped; available in source npz).

## Source

- [ms_academic_phy.npz](https://raw.githubusercontent.com/shchur/gnn-benchmark/master/data/npz/ms_academic_phy.npz) — 16,374,005 bytes, sha256 `e4d68468eba5fb8f…`, fetched 2026-07-23

**License**: MIT (shchur packaging)  
**Citation**: Shchur et al. Pitfalls of GNN Evaluation. 2018.

- https://github.com/shchur/gnn-benchmark

## Caveats

- 8,415-dim features exceed the CSV budget — label-only nodes.csv; features remain in the source npz.

---
*Generated by converters/make_cards.py; stats from metadata.json. Raw files: `source/`. NEExT tables: `neext/`.*