# Minesweeper

**Task**: Mine prediction on a synthetic grid  
**Size band**: medium · **Label type**: structural  
**Label column**: `is_mine` · **Converter**: `converters/convert_npz.py`

Synthetic 100x100 grid graph in the minesweeper style: 20% of cells are mines, features are one-hot counts of neighboring mines. Fully structural and regular — a controlled probe of what egonet embeddings can and cannot see.

## Converted graphs (neext/)

| graph | nodes | edges | classes | feature cols | isolated | class counts |
|---|---|---|---|---|---|---|
| default | 10,000 | 39,402 | 2 | 7 | 0 | 0: 8,000, 1: 2,000 |

*Conversion notes*: yandex heterophilous-graphs npz; edges symmetrized; 7 f_* feature columns.

## Source

- [minesweeper.npz](https://raw.githubusercontent.com/yandex-research/heterophilous-graphs/main/data/minesweeper.npz) — 135,045 bytes, sha256 `e664c8dacf1e8ac4…`, fetched 2026-07-23

**License**: MIT (yandex-research)  
**Citation**: Platonov et al. ICLR 2023.

- https://github.com/yandex-research/heterophilous-graphs

## Caveats

- Synthetic; the known optimal strategy bounds achievable AUC.

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