IsoSEL: Isometric Structural Entropy Learning for Deep Graph Clustering in Hyperbolic Space

April 15, 2025 ยท View on GitHub

Paper URL: https://arxiv.org/abs/2504.09970

The Extension of [ICML2024 (Oral)]: LSEnet: Lorentz Structural Entropy Neural Network for Deep Graph Clustering.

Get Started

Firstly, install all the required pakages and Python==3.9.0.

cd ./DSE_clustering
pip install -r requirements.txt

Run main.py to train and test model.

python main.py

You can design your own configurations in configs directory by .json files.

Pay attention to the following codes to load or save your configurations:

parser = argparse.ArgumentParser(description='Lorentz Structural Entropy')
# ...add_argument
configs = parser.parse_args()

# save to json file
with open(f'./configs/{configs.dataset}.json', 'wt') as f:
    json.dump(vars(configs), f, indent=4)

# load from json file
configs_dict = vars(configs)
with open(f'./configs/{configs.dataset}.json', 'rt') as f:
    configs_dict.update(json.load(f))
configs = DotDict(configs_dict)

f.close()

Model Architecture

football

Figure 1. In IsoSEL, we learn the partitioning tree of structural information with the LSEnet in hyperbolic space, and then propose Lorentz tree contrastive learning to refine the partitioning tree for graph clustering without K.

Main Experimental Results

football

Visualization

football
Figure 2. Visualization of real and predicted data.


football
Figure 2. Visualization of hyperbolic partitioning trees.