---
license: cc-by-4.0
task_categories:
- graph-ml
language:
- en
pretty_name: Property-Driven GNN Benchmark
size_categories:
- 1M<n<10M
tags:
- graph-neural-networks
- benchmark
- gnn-expressiveness
- formal-methods
- alloy
- graph-classification
---

# Property-Driven GNN Benchmark

A large-scale benchmark for evaluating the expressive power of Graph Neural Networks (GNNs) across 16 fundamental graph properties. This dataset accompanies the paper *"Systematic Property-Driven Evaluation of
GNN Expressiveness"*.

## Summary

The benchmark contains **352 graph-classification datasets** organized into two families:

- **GraphRandom** (176 datasets): graphs that either satisfy or randomly violate a given property.
- **GraphPerturb** (176 datasets): each positive graph is paired with a structurally similar negative counterpart that differs by only one or two edges.

All datasets were generated using **Alloy**, a relational-logic specification language and SAT-based analyzer, ensuring exhaustive enumeration of positive samples and verifiable negative samples.

## Quick Start

### Download the entire dataset

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="anonymousPaper5674/property-driven-gnn",
    repo_type="dataset",
    local_dir="./dataset",
)
```

### Load a single CSV file and reconstruct one graph

```python
import pandas as pd
import numpy as np

# Each row encodes one graph
df = pd.read_csv("./dataset/GraphRandom-Train/antisymmetric.csv", header=None)

row = df.iloc[0].values
n = int(np.sqrt(len(row) - 1))   # graph size (number of nodes)
adj = row[:-1].reshape(n, n)     # n x n adjacency matrix (directed)
label = int(row[-1])             # 1 if the graph satisfies the property, else 0

print(f"Graph with {n} nodes, label = {label}")
print(adj)
```

## Data Format

Each CSV file represents one dataset. The format is:

- **No header row.**
- **Each row encodes one graph.**
- The **first n² values** are the entries of the n×n adjacency matrix, flattened in **row-major order**: `adj[i][j] = row[i * n + j]`.
- The **last value** is the binary label: `1` if the graph satisfies the property, `0` otherwise.
- Edges are **directed**: `adj[i][j] = 1` means there is an edge from node *i* to node *j*.
- **Self-loops are allowed** (diagonal entries can be `1`).
- Graph size *n* can be inferred from the row length: `n = sqrt(len(row) - 1)`.

## Folder Structure

```text
Property-Driven-GNN/
├── GraphRandom-Train/
│   └── <property>.csv                   
├── GraphRandom-Test/
│   └── <property>/
│       └── <property>_new<k>.csv        
├── GraphPerturb-Train/
│   └── <property>.csv                     
└── GraphPerturb-Test/
    └── <property>/
        └── <property>_new<k>.csv         
```

## Code

The reference implementation, including data loading, training, and evaluation across nine global pooling methods, is available at:
`https://anonymous.4open.science/r/Property-Driven-GNN/README.md`