DCBA Data Set

This repository contains the DCBA data set along with the necessary code to generate and load the data. It is part of the codebase for the paper: "Graph Data Augmentation via Contrastive Generator Inversion (DCBA)" accepted to the 5th Learning on Graphs Conference (Boston, USA, 2026).

Runtime configuration

1. Install uv. uv manages the Python environment and dependency installation for this project.
2. From the repository root, sync the project environment and install dependencies:

   uv sync

3. Resolve Julia dependencies (required on first setup):

   uv run python -c "import juliapkg; juliapkg.resolve(force=True)"

4. Additionally, to use DVC with Google Drive as remote storage, install:

   uv tool install 'dvc[gdrive]'

5. Download the data required for experiments using DVC:

   dvc pull

6. Install pre-commit:

   uv run pre-commit install --config .pre-commit-config.yaml

   This tool will automatically check code formatting (it's a very convenient configuration) and run tests before each commit. To skip checks, use git commit --no-verify; to scan all files execute: uv run pre-commit run --all-files --config .pre-commit-config.yaml.

Datasets

All datasets live under data/ and are tracked by DVC. After dvc pull the following are available:

Dataset	Instances	Graphs	Notes
abcd-big	8 744	87 440	big set; used in the paper
abcd-borderline	486	4 860	hard set; used in the paper
abcd-interim	200	2 000	small set for test trainings
abcd-distinct-comms	200	1 999	easy set with separated communities

Graphs within abcd-big are prefixed with their chunk name in the instance ID (e.g. chunk-7/3f1a...). Instances that timed out mid-generation are excluded by default.

To download the dataset, one must authenticate with a Google account that has access to the shared Google Drive: https://drive.google.com/drive/u/1/folders/1YdaLLIRZNaptO6QzHpfS2IrfPoEFyvYq. If you need access, please contact the authors.

Data loaders

from dcba_data_set.graph_io import load_dataset, DCBAHeteroData

records = load_dataset("data/abcd-big")

graph: DCBAHeteroData = DCBAHeteroData.from_replica_record(
    records[0].replicas[0],
    instance_id=records[0].instance_id,
    net_type=records[0].net_type,
)

load_dataset handles both flat (report.json at root) and chunked layouts automatically. Pass discard_failed=False to retain instances that timed out with partial replica sets.

Generators

To run the code, execute: uv run dcba-data-set <path to the configuration file>. Example configs are provided in scripts/configs/example_generate/.

This repository provides three functionalities:

1. ABCD Generator

Generates a single ABCD graph from a given configuration. Set experiment_type: "generate-abcd" in the config. See scripts/configs/example_generate/abcd.yaml for an example.

2. mABCD Generator

Generates a single multilayer mABCD graph from a given configuration. Set experiment_type: "generate-mabcd" in the config. See scripts/configs/example_generate/mabcd.yaml for an example.

3. Dataset Generator

Samples multiple configurations from provided parameter ranges and generates a network for each. Set experiment_type: "generate-dataset" in the config. See scripts/configs/ for examples.

Citing the code

If you use the data set or the code, please consider citing us:

@inproceedings{stolarski2026dcba,
   title = {Graph Data Augmentation via Contrastive Generator Inversion (DCBA)},
   author = {
      Stolarski, Mateusz and Czuba, Micha{\l} and Krai\'{n}ski, \L{}ukasz and Musial, Katarzyna and
      Pra\l{}at, Pawe\l{} and Kami\'{n}ski, Bogumi\l{} and Br{\'o}dka, Piotr
   },
   booktitle = {Proceedings of the Fifth Learning on Graphs Conference},
   series = {Proceedings of Machine Learning Research},
   publisher = {PMLR},
   year = {2026},
   doi = {10.48550/arXiv.2610.05653},
}

Acknowledgement

This research was partially supported by: Horizon Europe, EU, grant no. 10108632; National Science Centre, Poland, grant no. 2022/45/B/ST6/04145; Polish National Agency for Academic Exchange, Strategic Partnerships, grant no. BPI/PST/2024/1/00129/U/00001; Wrocław University of Science and Technology, Academia Professorum Iuniorum and Minigrants projects. Views and opinions expressed here are those of the authors and do not necessarily reflect those of the funding agencies.