This is the pytorch implementation for our ICLR 2026 Oral:

Wei Ju, Siyu Yi, Kangjie Zheng, Yifan Wang, Ziyue Qiao, Li Shen, Yongdao Zhou, Xiaochun Cao, Jiancheng Lv. Compactness and Consistency: A Conjoint Framework for Deep Graph Clustering. ICLR 2026 Oral

---

🛠️Overview

CoCo is a deep graph clustering framework designed to learn compact and semantically consistent node representations from graph data. The method addresses two core issues in existing graph clustering methods:

• limited ability of local message passing to capture long-range dependencies;
• redundancy and noise in graph data that harm representation quality.

To solve this, CoCo combines:

1. Local-view and global-view feature extraction using graph convolutional filters on the original graph and a graph diffusion graph;
2. Compactness learning via GMM-based low-rank subspace training and feature reconstruction;
3. Consistency learning by aligning node-wise similarity distributions across the two views.

After training, the local and global compact embeddings are fused and K-means is applied to obtain final clustering assignments.

---

🚀 Quickstart

1.Environment

The paper implementation is based on PyTorch.

conda create -n coco python=3.10 -y
conda activate coco
pip install -r requirements.txt

---

2.Dataset preparation

The main graph clustering experiments use the following datasets:

• Cora
• AMAP
• BAT
• EAT
• UAT

Data layout:

dataset/<dataset>/
├── <dataset>_feat.npy
├── <dataset>_label.npy
└── <dataset>_adj.npy

---

3.RUN

Main clustering experiments

python train.py --dataset cora 

Dataset configuration could be found in setup.py.

The reported metrics are:

• ACC: clustering accuracy
• NMI: normalized mutual information
• ARI: adjusted rand index
• F1: macro F1-score

For reproducibility, report the mean ± standard deviation over 10 runs.

---

📈Main results

Graph clustering

Dataset	ACC	NMI	ARI	F1
Cora	79.36	60.71	58.76	77.95
AMAP	79.27	68.85	60.94	72.36
BAT	78.85	55.00	53.52	78.56
EAT	58.87	34.10	27.91	58.06
UAT	59.68	30.12	29.46	58.03

---

📝Citation

If you find this work useful, please cite:

@inproceedings{ju2026compactness,
  title={Compactness and Consistency: A Conjoint Framework for Deep Graph Clustering},
  author={Ju, Wei and Yi, Siyu and Zheng, Kangjie and Wang, Yifan and Qiao, Ziyue and Shen, Li and Zhou, Yongdao and Cao, Xiaochun and Lv, Jiancheng},
  booktitle={The Fourteenth International Conference on Learning Representations},
  year={2026}
}

---

🙏Acknowledgment

This repository utilizes the pre-processed datasets provided by the following sources. We gratefully acknowledge the authors and contributors for their valuable work and open sharing:

• yueliu1999/Awesome-Deep-Graph-Clustering