graphbench-lib 0.1.2.3

Next-generation graph learning benchmarking.

GraphBench: Next generation graph learning benchmarking

This is the package associated with the paper GraphBench: Next generation graph learning benchmarking.

It contains the code and tools necessary to load the datasets associated with the benchmark. GraphBench is a collection of benchmarking datasets across domains and tasks obtained from real world and synthetic applications.

Features

GraphBench comes as a Python package with the following features:

Data Loading	Efficiently loads graph datasets for benchmarking and experimentation across all domains and tasks.
Metric Evaluation	Supports a wide range of evaluation metrics for graph learning tasks.
Automated Model Tuning	Integrates SMAC3 for automatic hyperparameter optimization of user models.

Installation

We recommend using Anaconda/Miniconda during setup. The installation process is done in two steps:

1. Install swig via Anaconda/Miniconda (if using the optimization module): conda install swig

2. GraphBench can then be easily installed using the python package manager pip: pip install graphbench-lib

For using the optimzation module graphbench-lib[tuning] needs to be installed instead.

Please make sure to update the installation of GraphBench before running the benchmark for best results. Alternatively one can also install from source:

git clone https://github.com/graphbench/package
cd package
pip install -e . 

Dependencies

GraphBench uses some python packages to enable loading and processing of datasets. An overview of the required packages can be found in the list below:

• torch
• torch_geometric
• networkx
• torchmetrics
• numpy
• pandas
• requests
• tqdm
• scikit-learn

Optionally smac3 is used for the optimization module.

Usage

The package can be easily used to get selected datasets from the GraphBench tasks:

import graphbench
Loader = graphbench.Loader(root, dataset_name)
datasets = Loader.load()

Furthermore, standardized evaluation metrics can be obtained using the following methods:

Evaluator = graphbench.Evaluator(metric_name)
metric_results = Evaluator.evaluate()

In order to use all datasets of a domain easily, each domain corresponds to one dataset_name variable:

Domain	Dataset_name
Social media	socialnetwork
Combinatorial optimization	co
SAT solving	sat
Algorithmic reasoning	algorithmic_reasoning_easy, algorithmic_reasoning_medium, algorithmic_reasoning_hard
Electronic circuits	electronic_circuits
Chip design	chipdesign
Weather forecasting	weather

Note, that for algorithmic reasoning the download always includes all datasets for a given task. We plan to change this in the future.

For a full list of the datasets, see the accompanying website or the datasets.csv file. The corresponding metrics can be found in the master.csv file.

Citing GraphBench:

If you use GraphBench or GraphBench datasets in your work please cite our paper:

@article{GraphBench,
title={GraphBench: Next-generation graph learning benchmarking}, 
author={Timo Stoll and Chendi Qian and Ben Finkelshtein and Ali Parviz and Darius Weber and Fabrizio Frasca and Hadar Shavit and Antoine Siraudin and Arman Mielke and Marie Anastacio and Erik Müller and Maya Bechler-Speicher and Michael Bronstein and Mikhail Galkin and Holger Hoos and Mathias Niepert and Bryan Perozzi and Jan Tönshoff and Christopher Morris},
year={2025},
journal={arXiv preprint arXiv:2512.04475}
}