Network_datasets / README.md
huzhzh's picture
Document the sharded layout
b874814 verified
|
Raw History Blame Contribute Delete
3.03 kB
---
license: other
pretty_name: Network datasets for BiT-HyRL
task_categories:
- graph-ml
tags:
- network-science
- complex-networks
- graph
- resilience
size_categories:
- 10K<n<100K
---
# Network datasets
Topologies used in *Bimodal Topology and Hybrid Reinforcement Learning for Resilient
Distributed Control Systems Against Cascading Failures* (Reliability Engineering & System
Safety, JRESS-D-26-02993).
## Contents
```text
dataset/testdata/ 30 evaluation topologies (GML) and the raw SNAP as-733 edge lists
dataset/all/real/ 231 real topologies (GML)
dataset/all/syn/<family>/ 23 526 synthetic topologies (GML), one directory per family
gdm/gdm_nx_best.pth pretrained GDM model used by the GDM attack
```
Every graph is a simple undirected graph: nodes are relabelled `0..n-1`, self-loops and
parallel edges are removed, and the edge list is stored in GML.
The synthetic corpus is split by generator family because the hub allows at most 10 000
files per directory. The training scripts in the code repository read a flat
`dataset/all/syn/`, so flatten the tree after downloading:
```bash
find dataset/all/syn -mindepth 2 -name "*.gml" -exec mv -n {} dataset/all/syn/ \;
find dataset/all/syn -mindepth 1 -type d -empty -delete
```
## Synthetic topologies
File names follow `<family>_n<N>_i<K>.gml`, with `N` the node count and `K` the draw index.
Watts-Strogatz files also carry the rewiring parameters, for example `WS_n100_k6_p03_i0.gml`.
| Family | Count | Generator |
|---|---|---|
| `ER` | 3 825 | Erdős-Rényi |
| `BA` | 3 811 | Barabási-Albert preferential attachment |
| `WS` | 3 808 | Watts-Strogatz small world |
| `WSlow`, `WShigh` | 2 400 each | Watts-Strogatz at low and high mean degree |
| `RG` | 2 400 | random geometric |
| `HK` | 2 400 | Holme-Kim |
| `BAdense` | 2 400 | dense Barabási-Albert |
| `planted`, `Erdos` | 11 and 10 | planted partition and Erdős-Rényi benchmarks |
The corpus is regenerated with `scripts/generate_training_data.py` in the code repository.
## Real topologies
SNAP `as-733` BGP-AS snapshots, Internet Topology Zoo ISP backbones, and networks assembled
for this study. The raw SNAP edge lists are in `dataset/testdata/AS-*.txt`.
## Evaluation set
`dataset/testdata/` holds the thirty graphs behind every table and figure, together with the
raw SNAP as-733 edge lists they were built from. `n`, `m` and the SHA-256 of the sorted edge
list of each graph are pinned in `configs/graphs.json` of the code repository.
## GDM weights
`gdm/gdm_nx_best.pth` is the model loaded by
`network_dismantling/GDM/python_interface.py`, which reads the `models_newpg/` directory of
that package. Copy the file there before running the GDM attack.
## Code
<https://github.com/kuaiii/BiT_HyRL>
## Licence
The synthetic graphs were generated for this study and may be used freely with attribution.
The real topologies are redistributed from their original public sources (SNAP, the Internet
Topology Zoo) and keep the terms of those sources.