HaiFormer-6G: Hierarchical AI-Native Transformer–GNN Framework for Autonomous 6G Networks

A Research Framework for AI-Native Autonomous 6G Network Management

Graph Neural Networks • Spatio-Temporal Transformers • Digital Twin • Explainable AI • Multi-Task Learning • Uncertainty Estimation

---

Overview

HaiFormer-6G is a research-oriented deep learning framework designed for AI-native autonomous 6G network management.

Unlike conventional approaches that separately employ Graph Neural Networks (GNNs), Transformers, or Digital Twins, HaiFormer-6G integrates these components into a closed-loop hierarchical intelligence engine capable of continuously understanding, predicting, explaining, and adapting to rapidly changing network conditions.

The framework models the communication network as a dynamic graph whose topology evolves with real-time telemetry, enabling intelligent resource management, traffic prediction, anomaly detection, and autonomous decision-making.

Research Motivation & Question

Future 6G networks are expected to operate under strict performance constraints requiring massive topology dynamics, AI-native control, zero-touch management, self-optimization, self-healing, and explainable decision-making.

While existing literature features isolated applications of Digital Twins, Graph Neural Networks, or Transformer architectures, most models target singular tasks or lack architectural integration, leaving open challenges in continual adaptation, scalability, explainability, uncertainty estimation, autonomous operation, and multi-task intelligence.

Research Question: How can an AI-native 6G network autonomously adapt to rapidly changing network conditions using a hierarchical Graph–Transformer architecture while providing explainable and confidence-aware decisions?

---

Main Contributions

Primary Contribution

• Hierarchical AI-Native Intelligence Engine that unifies non-Euclidean spatial learning, temporal dependency modeling, topology evolution, and intelligent network management within a single optimization framework.

Supporting Contributions

• Dynamic Graph Generation from live network telemetry.

• Cross-Attention Fusion between graph and temporal representations.

• Heterogeneous Multi-Task Learning for:

  • Traffic Matrix Prediction
  • QoS Jitter Estimation
  • Network Anomaly Detection
  • Bottleneck Detection

• Confidence-Aware Prediction using uncertainty estimation.

• Built-in Explainable AI (XAI) for interpretable decision making.

• Closed-Loop Digital Twin interaction for continuous network adaptation.

---

Framework Architecture

                 Live 6G Network
                        │
                        ▼
        Dynamic Network Telemetry Stream
                        │
                        ▼
            Dynamic Graph Constructor
                        │
                        ▼
      Spatial Graph Neural Network Encoder
                        │
                        ▼
      Hierarchical Transformer Encoder
                        │
                        ▼
         Cross-Attention Fusion Module
                        │
                        ▼
        Shared Latent Representation
                        │
 ┌──────────────┬──────────────┬──────────────┬──────────────┐
 │ Traffic      │ QoS          │ Anomaly      │ Bottleneck   │
 │ Prediction   │ Prediction   │ Detection    │ Detection    │
 └──────────────┴──────────────┴──────────────┴──────────────┘
                        │
                        ▼
          Uncertainty Estimation Layer
                        │
                        ▼
          Explainability Module (XAI)
                        │
                        ▼
      Autonomous Decision Engine
                        │
                        ▼
      Closed-Loop Digital Twin Feedback

---

Repository Structure

HaiFormer-6G/
│
├── model.ipynb
├── geant
├── best_haiformer6g_static_aligned_60ep.pt  
├── academic_outputs_png/
│   ├── attention_heatmap.png
│   ├── confusion_matrix.png
│   ├── residual_distribution.png
│   ├── accuracy_curve.png
│   └── traffic_prediction.png
├── README.md
└── LICENSE

---

Core Features

• Hierarchical Graph–Transformer Architecture
• Dynamic Graph Learning
• Multi-Task Deep Learning
• Cross-Attention Fusion
• Digital Twin Integration
• Explainable AI
• Confidence-Aware Prediction
• Publication-Quality Visualization
• End-to-End Research Pipeline

---

Installation

git clone https://github.com/mahins32/HaiFormer-6G.git

cd HaiFormer-6G

pip install -r requirements.txt

---

Running the Framework

jupyter notebook notebooks/main_pipeline.ipynb

Run the notebook sequentially to reproduce:

1. Dataset Preparation
2. Graph Construction
3. Model Training
4. Evaluation
5. Visualization
6. Publication Figures

---

Dataset

The framework is developed using the GÉANT backbone network traffic dataset, represented as dynamic graph sequences for spatio-temporal learning.

---

Future Research Directions

• Online Continual Learning
• Federated AI-Native 6G
• Reinforcement Learning Integration
• Large Network Foundation Models
• Self-Evolving Digital Twins
• Multi-Agent Network Intelligence

---

Citation

@misc{haiformer6g2026,
  title={{HaiFormer-6G}: Hierarchical {AI}-Native Transformer-{GNN} Framework for Autonomous {6G} Networks},
  author={Mahin, Saiful Islam},
  year={2026},
  note={Open-Source Research Repository},
  howpublished={https://github.com/mahins32/HaiFormer-6G}
}

---

License

Released under the MIT License.

---

👤 Author

Saiful Islam Mahin

Department of Electrical and Electronic Engineering

University of Asia Pacific

Email: mahinss399@gmail.com