DiffGraph
[WSDM'2025] "DiffGraph: Heterogeneous Graph Diffusion Model"
Install / Use
/learn @HKUDS/DiffGraphREADME
🌌 DiffGraph: Heterogeneous Graph Diffusion Model
<img src="https://img.shields.io/badge/WSDM-2025-FF6B6B?style=for-the-badge&logo=ieee&logoColor=white" alt="WSDM 2025"> <img src="https://img.shields.io/badge/PyTorch-1.12.1-EE4C2C?style=for-the-badge&logo=pytorch&logoColor=white" alt="PyTorch"> <img src="https://img.shields.io/badge/DGL-1.0.2-00C7B7?style=for-the-badge&logo=dgl&logoColor=white" alt="DGL"> <img src="https://img.shields.io/badge/Python-3.8-3776AB?style=for-the-badge&logo=python&logoColor=white" alt="Python"> <img src="https://readme-typing-svg.herokuapp.com?font=Fira+Code&size=22&duration=3000&pause=1000&color=00D4AA&background=FFFFFF00¢er=true&vCenter=true&width=600&lines=%F0%9F%9A%80+Latent+Diffusion+for+Graphs;%E2%9A%A1+Noise-Resilient+Learning;%F0%9F%8C%8A+Cross-View+Semantic+Fusion" alt="Typing SVG" /> <div align="center"> <table align="center"> <tr> <td align="center">██████╗ ██╗███████╗███████╗ ██████╗ ██████╗ █████╗ ██████╗ ██╗ ██╗
██╔══██╗██║██╔════╝██╔════╝██╔════╝ ██╔══██╗██╔══██╗██╔══██╗██║ ██║
██║ ██║██║█████╗ █████╗ ██║ ███╗██████╔╝███████║██████╔╝███████║
██║ ██║██║██╔══╝ ██╔══╝ ██║ ██║██╔══██╗██╔══██║██╔═══╝ ██╔══██║
██████╔╝██║██║ ██║ ╚██████╔╝██║ ██║██║ ██║██║ ██║ ██║
╚═════╝ ╚═╝╚═╝ ╚═╝ ╚═════╝ ╚═╝ ╚═╝╚═╝ ╚═╝╚═╝ ╚═╝ ╚═╝
🌟 Advancing Heterogeneous Graph Intelligence through Novel Latent Diffusion Strategies
🎯 Mission Statement
</div>"In the labyrinth of heterogeneous data, where noise corrupts truth and complexity obscures patterns, DiffGraph emerges as the quantum leap in graph intelligence - wielding the power of latent diffusion to transform chaos into clarity."
🧠 Neural Architecture Overview
<div align="center"> <img src="./HDL.jpg" alt="DiffGraph Architecture" width="85%"> <br> <em>🔬 The Heterogeneous Graph Diffusion Pipeline: From Noisy Reality to Pure Intelligence</em> </div>🌟 Core Innovation Matrix
| 🔥 Component | 🎮 Technology | 🎯 Breakthrough | |------------------|-------------------|---------------------| | Latent Diffusion Engine | Gaussian Noise Injection + Progressive Denoising | Eliminates heterogeneous noise while preserving semantic integrity | | Cross-View Semantic Fusion | Auxiliary-to-Target Graph Transformation | Maximizes mutual information across graph modalities | | Quantum GCN Layers | Multi-relational Message Passing | Captures complex heterogeneous transitions | | Neural Denoising Network | Time-Conditioned MLP Architecture | Reconstructs pure graph representations |
🚀 Performance Overview
<div align="center">📊 Main Results Summary
| Task | Dataset | Best Baseline | DiffGraph | Improvement | |----------|-------------|-------------------|---------------|-----------------| | Link Prediction | Tmall | 0.0463 (R@20) | 0.0589 | +27.21% ⚡ | | | Retail Rocket | 0.0524 (R@20) | 0.0620 | +18.32% 🚀 | | | IJCAI | 0.0136 (R@20) | 0.0171 | +25.74% 💎 | | Node Classification | DBLP | 91.97% (Micro-F1) | 93.81% | +2.00% 📈 | | | AMiner | 82.46% (Micro-F1) | 83.29% | +1.01% 🎯 | | | Industry | 79.82% (AUC) | 80.25% | +0.54% 💪 |
</div>📈 Detailed Experimental Analysis
<details> <summary><b>🔍 Click to expand detailed results</b></summary>📊 Link Prediction - Complete Results
<div align="center">| Dataset | Metric | MATN | HGT | MBGCN | DiffGraph | Gain | |-------------|------------|----------|---------|-----------|---------------|----------| | Tmall | Recall@20 | 0.0463 | 0.0431 | 0.0419 | 0.0589 | +27.21% | | | NDCG@20 | 0.0197 | 0.0192 | 0.0179 | 0.0274 | +39.09% | | Retail Rocket | Recall@20 | 0.0524 | 0.0413 | 0.0492 | 0.0620 | +18.32% | | | NDCG@20 | 0.0302 | 0.0250 | 0.0258 | 0.0367 | +21.52% | | IJCAI | Recall@20 | 0.0136 | 0.0126 | 0.0112 | 0.0171 | +25.74% | | | NDCG@20 | 0.0054 | 0.0051 | 0.0045 | 0.0063 | +16.67% |
</div>🎯 Node Classification - Best Results
<div align="center">| Dataset | Setting | Best Baseline | DiffGraph | Metric | |-------------|-------------|-------------------|---------------|------------| | DBLP | 60 per class | HeCo: 91.59±0.2 | 93.81±0.3 | Micro-F1 | | | 60 per class | HeCo: 98.59±0.1 | 99.21±0.1 | AUC | | AMiner | 40 per class | HeCo: 80.53±0.7 | 83.29±1.3 | Micro-F1 | | | 40 per class | HeCo: 92.11±0.6 | 94.41±0.8 | AUC | | Industry | Full dataset | HGT: 0.7982 | 0.8025 | AUC |
</div> </details>🏗️ System Architecture
🌌 DiffGraph Neural Framework
├── 🔥 DiffGraph-Rec/ # Link Prediction Engine
│ ├── 🧠 Model.py # Core HGDM Implementation
│ ├── 📊 DataHandler.py # Multi-behavior Data Processing
│ ├── ⚙️ main.py # Training & Evaluation Pipeline
│ ├── 🎛️ params.py # Hyperparameter Configuration
│ ├── 🗂️ data/ # Heterogeneous Datasets
│ │ ├── tmall/ # E-commerce Multi-behavior
│ │ ├── retail_rocket/ # Transaction Networks
│ │ └── ijcai_15/ # Competition Benchmark
│ └── 🛠️ Utils/ # Neural Utilities
├── 🎯 DiffGraph_NC/ # Node Classification Engine
│ ├── 🧠 Model.py # Academic Network HGDM
│ ├── 📊 DataHandler.py # Citation Network Processing
│ ├── ⚙️ main.py # Classification Pipeline
│ ├── 🎛️ params.py # Configuration Matrix
│ ├── 🗂️ data/ # Academic Datasets
│ │ ├── dblp/ # Database & AI Publications
│ │ └── aminer/ # Research Network
│ └── 🛠️ Utils/ # Classification Tools
└── 📖 README.md # This Neural Manual
🔬 Scientific Foundation
📜 Mathematical Formulation
Latent Heterogeneous Graph Diffusion Process:
𝒢ₛ* ↭^π 𝐄ₛ* →^φ 𝐄̃ₛ* →^φ' 𝐄̃ₛ* ↭^π' 𝒢̃ₛ*
Forward Diffusion Trajectory:
q(ℋₜ | ℋₜ₋₁) = 𝒩(ℋₜ; √(1-βₜ)ℋₜ₋₁, βₜ𝐈)
Reverse Denoising Process:
p(ℋₜ₋₁ | ℋₜ) = 𝒩(ℋₜ₋₁; μθ(ℋₜ,t), Σθ(ℋₜ,t))
🎯 Core Contributions
- 🌟 Latent Space Revolution: First heterogeneous graph diffusion in latent space, solving discrete graph generation challenges
- 🔄 Cross-View Intelligence: Novel auxiliary-to-target semantic transformation mechanism
- 🛡️ Noise Resilience: Superior robustness against heterogeneous data corruption
- ⚡ Scalable Architecture: Linear complexity with heterogeneous relation types
📊 Datasets & Benchmarks
<div align="center">| Task | Dataset | Scale | Domain | |----------|-------------|-----------|------------| | Link Prediction | Tmall | 31K users, 31K items | E-commerce Multi-behavior | | | Retail Rocket | 2K users, 30K items | Transaction Networks | | | IJCAI-15 | 17K users, 36K items | Competition Benchmark | | Node Classification | DBLP | 26K nodes, 4 classes | Academic Publications | | | AMiner | 56K nodes, 4 classes | Research Networks | | | Industry | 2M+ users | Gaming Platform |
Complete dataset details available in paper appendix
</div>🔬 Component Analysis
<div align="center">| Analysis Type | Key Finding | Performance Impact | |-------------------|-----------------|------------------------| | 🧩 Ablation Study | Diffusion module crucial | -11.0% without diffusion | | ⚙️ Hyperparameters | Optimal: 64-dim, 3-layers | Best at moderate complexity | | 🛡️ Noise Robustness | Superior resilience | 50% less degradation vs baselines | | ⚡ Efficiency | 2.6x faster training | Computational advantage | | 📊 Data Sparsity | Consistent gains | +31.4% on sparse data |
</div> <details> <summary><b>📊 Click to view detailed analysis</b></summary>🧩 Ablation Study
| Variant | Description | Tmall R@20 | Change | |-------------|-----------------|----------------|------------| | DiffGraph | Full model | 0.0589 | - | | -D | Remove diffusion | 0.0524 | -11.0% | | -H | Remove heterogeneous | 0.0463 | -21.4% | | DAE | Replace w/ autoencoder | 0.0531 | -9.8% |
🛡️ Noise Robustness (50% Noise)
| Behavior | DiffGraph Retention | HGT Retention | |--------------|-------------------------|-------------------| | Page View | 97.42% | 95.59% | | Favorite | 98.62% | 97.22% | | Cart | 96.73% | 95.82% |
📊 Data Sparsity Impact
- Sparse Users (< 8 interactions): +31.4% improvement
- Medium Users (< 35 interactions): +25.1% improvement
- Active Users (< 120 interactions): +19.4% improvement
🏆 Competitive Analysis
<div align="center">🎯 Performance Advantage
| **C
