tags:
- graph-neural-networks
- unsupervised-learning
- graph-clustering
- pytorch
🔍 Overview
This repository provides the official implementation of SE-GCL, an unsupervised graph clustering framework based on structural entropy and probabilistic optimization.
Unsupervised graph clustering aims to uncover latent community structures in graph data. However, existing methods often struggle to simultaneously capture local structural variations and global topological patterns, while deep learning approaches typically suffer from large parameter sizes and computational redundancy.
To address these challenges, SE-GCL:
Introduces a Structural Entropy Guidance (SEO) module to model node uncertainty and structural contribution
Designs a parameter-free optimization strategy to handle the non-differentiability of structural entropy
Employs a Clustering Search Module (CSM) with probabilistic updates for adaptive clustering
Eliminates the need for predefining the number of clusters
🚀 Features
✅ Unsupervised graph clustering
✅ No need to predefine cluster number
✅ Lightweight model with fewer parameters
✅ Captures both local and global structure
✅ Stable probabilistic clustering process
📊 Performance
SE-GCL achieves:
🔹 80%–96% performance of state-of-the-art methods
🔹 Using only 12%–55% model parameters
▶️ Usage
Run the main script:
python main.py
