DHG-Bench

DHG-Bench is a unified library for Deep Hypergraph Learning (DHGL), built on PyTorch and PyTorch Geometric . It integrates 17 state-of-the-art hypergraph neural network (HNN) algorithms and 22 hypergraph datasets with diverse characteristics.

📦 Installation

Follow the steps below to install and configure DHG-Bench properly for your local environment.

Required Dependencies:

DHG-Bench needs the following requirements to be satisfied beforehand:

• Python>=3.9.21
• Pytorch>=2.2.2
• torch_geometric>=2.6.1
• torch-cluster>=1.6.3
• torch-scatter>=2.1.2
• torch-sparse>=0.6.18
• torch-spline-conv>=1.2.2
• deeprobust==0.2.11
• ipdb==0.13.13
• numpy==1.24.3

Installation with pip [Recommended]

pip install dhg-bench

Installation for local development

# download the resporitary
cd dhgbench
# install required dependencies

Download Datasets

We include benchmark datasets in the data.zip archive. Users can simply extract the archive with

unzip data.zip

to obtain the data directory. The project structure should look like the following:

DHG-Bench
  ├── data
  │   ├── fair_data
  │   ├── hete_data
  |   ├── hgcls_data
  │   ├── trad_data
  |   └── ...
  └── dhgbench
  │   ├── lib_dataset
  │   ├── lib_models
  |   ├── lib_utils
  │   ├── lib_yamls
  │   ├── parameter_parser.py
  └── └── ...

🚀 Quick Start

The following demonstrates show how to quickly run HGNN algorithm on the Cora dataset for the node classification task.

Step 1: Import Package

from dhgbench.parameter_parser import parameter_parser,method_config,set_task_args
from dhgbench.lib_utils.exp_agent import ExpAgent
from dhgbench.lib_models.HNN.preprocessing import algo_preprocessing
from dhgbench.lib_dataset.data_base import HyperDataset
from dhgbench.lib_dataset.preprocessing import data_processing

Step 2: Load and Set Up Configuration

args = parameter_parser() 
args.dname, args.method, args.task_type, args.is_default = 'cora', 'HGNN', 'node_cls', False
args = method_config(args) 
args = set_task_args(args) 

Note that The is_default parameter indicates whether to use the model’s default configuration. If set to False, the model will instead load parameter settings specific to the given dataset. All model parameter files are provided in the lib_yamls directory.

Step 3: Load and Preprocess Dataset

data=HyperDataset(args) 

if args.task_type == 'hg_cls':
    data = data.multi_hypergraphs 
else:
    data = data_processing(args,data)
    data._initialization_()
    data = algo_preprocessing(data,args)

Step 4: Training and Evaluation

agent = ExpAgent(args)
agent.running(args.task_type,data)

⚙️ Configuration Argument Options

You can flexibly configure experiments with the following key parameters:

task_type

Specifies the type of task:

'node_cls', 'edge_pred', or 'hg_cls'

dname

Supported datasets include:

Node-level datasets:

'cora', 'pubmed', 'coauthor_cora','coauthor_dblp', 'ModelNet40', 'zoo', 'yelp', 'walmart-trips-100', 'trivago', 'actor','amazon', 'pokec', 'twitch', 'german', 'bail', 'credit'

Edge-level datasets:

'cora', 'pubmed', 'coauthor_cora', 'coauthor_dblp', 'actor', 'pokec'

Graph-level datasets:

'RHG_3', 'RHG_10', 'IMDB_dir_form', 'IMDB_dir_genre', 'stream_player', 'twitter_friend'

method

The algorithm to run. Supported algorithms include:

• Spectral-based methods:

  'HGNN','HyperGCN','HCHA','LEGCN','HyperND','HJRL','SheafHyperGNN','PhenomNN','DPHGNN','TFHNN'
  

• Spatial-based methods:

  'HNHN', 'UniGCNII', 'AllSetformer', 'EDHNN', 'HyperGT'
  

• Tensor-based methods:

  'EHNN', 'TMPHN'
  

You can also manually modify the configuration files located in lib_yamls directory for fine-grained control over hyperparameters.