ReHGNN: Reliability-Aware Heterogeneous Hypergraph Neural Recommender

Overview

ReHGNN is a deep learning framework for intelligent server recommendation in distributed large-model deployment scenarios. As large-scale AI models increasingly rely on distributed infrastructure for reliable service delivery, selecting optimal server combinations becomes critical for balancing deployment quality, system reliability, and service latency. Traditional rule-based or heuristic methods struggle with the inherent heterogeneity of system entities (users, models, servers) and the high-order deployment relationships (e.g., one model deployed on multiple servers, one user invoking multiple models).

This project addresses these challenges by modeling the deployment problem as a heterogeneous hypergraph neural network task, where:

• Heterogeneous nodes represent users, models, and servers with distinct feature spaces
• Hyperedges explicitly capture group-level relationships (user-model interactions, model-server deployments, server topology)
• Weakly-supervised learning leverages historical reliable deployment records to implicitly optimize multiple system objectives

The framework achieves superior performance in deployment recommendation accuracy and exhibits strong generalization across diverse data distributions.

---

Key Features

• Unified Heterogeneous Modeling: Seamlessly integrates multi-type entities (users, models, servers) into a single graph structure
• High-Order Relationship Encoding: Uses hypergraph structure to natively express "one-to-many" and "many-to-many" deployment patterns
• Implicit Multi-Objective Optimization: Learns from historical reliable deployments without explicit multi-objective weighting
• Scalable Architecture: Sparse hypergraph computations enable efficient training on large-scale systems
• Robust Generalization: Maintains stable performance across different regional focus distributions

---

Quick Start

Prerequisites

• Python 3.8+
• PyTorch 1.12+
• CUDA 11.0+ (for GPU acceleration)
• 16GB+ RAM (32GB+ recommended)

Installation

# Clone the repository
git clone https://github.com/your-organization/ReHGNN.git
cd ReHGNN

# Install dependencies
pip install -r requirements.txt

Minimal Working Example

1. Prepare Data Directory Structure

Organize your data as follows (use placeholder datasets for testing):

datasets/
├── <dataset_name>/
│   ├── user-train.csv          # User features (Lo, La, ServerID, W1, W2, Size)
│   ├── user-test.csv
│   ├── model.csv               # Model features (ModelType, ArenaScore, Modelsize, Modelresource)
│   ├── server.csv              # Server features (Lo, La, LinkBandwidth, ComputationCapacity, StorageCapacity)
│   ├── user-model-train.csv    # User-model interactions (wide format)
│   ├── user-model-test.csv
│   └── server topology.csv     # Server network topology (adjacency matrix)

2. Configure Basic Settings

Edit config/config.yaml to specify data paths and basic hyperparameters:

# Data paths (update to your actual paths)
data_root: 'datasets/<dataset_name>'
result_root: 'results'

# Model architecture (adjust based on your data scale)
k_positive: 5-20              # Number of positive samples (deployment redundancy)
n_hid: 64-256                 # Hidden layer dimension
dropout: 0.0-0.2              # Dropout rate

# Training settings
lr: 1e-4 to 1e-3              # Learning rate
max_epochs: 100-500           # Maximum training epochs
eval_k_list: [1, 3, 5, 10, 20]  # Evaluation K values for metrics

3. Train the Model

# Basic training with default configuration
python train.py

# Advanced: Specify custom configuration
python train.py --config config/custom_config.yaml

4. Run Inference

# Perform inference on test set
python inference_model_placement.py --checkpoint results/<experiment_dir>/checkpoints/best_model.pth

# Compare with baseline methods
python inference_model_placement.py --checkpoint <model_path> --compare_baselines

Obtaining Data

For research purposes, you may:

1. Use public benchmark datasets (e.g., adapt datasets from system/network research)
2. Contact the authors for collaboration opportunities (see Contact section)

Pre-trained Models

Pre-trained model checkpoints are not publicly released to protect proprietary experimental configurations. However, you can train models from scratch following the Quick Start guide. Typical training time: 4-7 hours on a single RTX 3070 GPU for datasets with ~10K users, ~200 models, ~1.5K servers.

---

Reproducing Experiments

Environment & Compatibility

Tested Environments

Component	Version	Notes
Python	3.8, 3.9, 3.10	3.8 recommended
PyTorch	1.12, 1.13, 2.0	GPU support required
CUDA	11.0, 11.3, 11.7	Match with PyTorch version
NumPy	1.21+	For numerical operations
Pandas	1.3+	For data loading
scikit-learn	1.0+	For t-SNE visualization
Matplotlib	3.5+	For plotting

Contact & Support

For questions, bug reports, or collaboration inquiries:

• Email: [gymorsiback@tju.edu.cn]

---

Acknowledgments

Special thanks to the open-source community for providing foundational tools:

• PyTorch for deep learning framework
• DeepHypergraph (DHG) for hypergraph utilities
• scikit-learn for evaluation metrics

---

Last Updated: November 2025
Version: 1.0.6