QGHNN: A Quantum Graph Hamiltonian Neural Network

Wenxuan Wang, Jinjing Shi, Nan Guan, Shichao Zhang, Xuelong Li

TL;DR

This paper introduces the Quantum Graph Hamiltonian Neural Network (QGHNN), a novel approach that bridges graph representation learning with quantum Hamiltonian mechanics. By mapping graph structures to topological quantum systems, QGHNN aims to enhance graph learning capabilities on noisy intermediate-scale quantum (NISQ) computers, demonstrating superior performance and robustness against noise.

What Problem Does This Paper Address?

The paper tackles the challenge of effectively representing and learning from graph-structured data, which is prevalent in many real-world applications but difficult for traditional machine learning models. While Graph Neural Networks (GNNs) are designed for this, and Quantum Neural Networks (QNNs) offer potential for quantum parallelism, current QNN research often overlooks the crucial link between quantum state encoding and graph structures. This disconnect limits the full exploitation of quantum advantages. The authors aim to bridge this gap by developing a QNN that leverages quantum Hamiltonian principles to better capture graph information, especially on NISQ devices which are prone to noise.

Methods

The core innovation lies in the Quantum Graph Hamiltonian Learning (QGHL) method and the subsequent Quantum Graph Hamiltonian Neural Network (QGHNN).

  • QGHL: This method establishes a mapping between a graph ($G$) and the Hamiltonian of a topological quantum system ($H_m$). It utilizes the lattice properties of topological quantum systems to construct parameterized quantum circuits. These circuits evolve an initial quantum state to a final state that encodes the graph's information. The process involves optimizing circuit parameters by minimizing a loss function that quantifies the difference between the evolved state and the target Hamiltonian.
  • QGHNN: Built upon QGHL, QGHNN uses these parameterized quantum circuits as its "neuron units." It trains these circuits by minimizing a loss function derived from the graph's Hamiltonian and the quantum states. The parameters are updated using gradient descent. The framework consists of input data (graph and its mapped Hamiltonian), quantum neuron units (parameterized circuits), loss function computation, and parameter updating.

Datasets

The paper does not introduce new datasets. Instead, it uses synthetic graph data for experiments, specifically focusing on learning target graphs ($G_t$) of varying sizes (4, 5, and 6 qubits).

Key Insights and Ideas

  • Bridging Graphs and Hamiltonians: The fundamental insight is that the Hamiltonian of a topological quantum system can serve as a powerful representation for graph structures. This allows quantum systems to "learn" graph properties by learning their corresponding Hamiltonians.
  • NISQ-Friendly Approach: QGHNN is designed with NISQ devices in mind, aiming to be robust against quantum noise. The paper suggests that its architecture inherently reduces the impact of noise, a critical factor for practical quantum machine learning.
  • Unified Framework: QGHNN provides a unified framework that combines the strengths of GNNs (graph representation) with QNNs (quantum parallelism and potential for noise resilience) by leveraging Hamiltonian learning.
  • Parameter Efficiency: The results suggest that QGHNN can achieve accurate graph verification with fewer parameters compared to other quantum machine learning models.

Key Results

Experiments were conducted on the PennyLane quantum platform using 4, 5, and 6 qubits.

  • Superior Performance: QGHNN consistently outperformed other quantum machine learning models (VQE, QAOA, QNN) across all assessment metrics (Mean Squared Error, Cosine Similarity, Frobenius Norm, Correlation Coefficient).
  • High Accuracy: With 4 qubits, QGHNN achieved a mean squared error of 0.004 and a cosine similarity of 99.8%.
  • Robustness: The model demonstrated high robustness against quantum noise, maintaining strong performance even with an increasing number of qubits. The paper attributes this to improved noise management.
  • Loss Function Convergence: The loss function consistently converged towards -1.000 across all experiments, indicating good learning capacity and stability.

Limitations and Future Work

  • Scalability: While QGHNN shows promise, the paper acknowledges that simulating QGHL and QGHNN on classical systems becomes exponentially demanding with more qubits. This highlights the need for actual quantum hardware for larger-scale applications.
  • Noise Impact: Although QGHNN is designed to be noise-resilient, the paper notes that accuracy slightly decreases with an increasing number of qubits, suggesting that quantum noise still has an impact, albeit managed.
  • Future Applications: The authors suggest that QGHNN has significant potential for future applications in quantum knowledge graphs and recommendation systems, and for broader integration with deep learning models. Further research could explore more complex graph structures and real-world datasets.

Summary updated

Abstract

Representing and learning from graphs is essential for developing effective machine learning models tailored to non-Euclidean data. While Graph Neural Networks (GNNs) strive to address the challenges posed by complex, high-dimensional graph data, Quantum Neural Networks (QNNs) present a compelling alternative due to their potential for quantum parallelism. However, much of the current QNN research tends to overlook the vital connection between quantum state encoding and graph structures, which limits the full exploitation of quantum computational advantages. To address these challenges, this paper introduces a quantum graph Hamiltonian neural network (QGHNN) to enhance graph representation and learning on noisy intermediate-scale quantum computers. Concretely, a quantum graph Hamiltonian learning method (QGHL) is first created by mapping graphs to the Hamiltonian of the topological quantum system. Then, QGHNN based on QGHL is presented, which trains parameters by minimizing the loss function and uses the gradient descent method to learn the graph. Experiments on the PennyLane quantum platform reveal that QGHNN outperforms all assessment metrics, achieving the lowest mean squared error of <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">0.004</b> and the maximum cosine similarity of <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">99.8%</b>, which shows that QGHNN not only excels in representing and learning graph information, but it also has high robustness ability. QGHNN can reduce the impact of quantum channel noise and has significant potential application in future research of quantum knowledge graphs and recommendation systems.
BibTeX
@article{wang2025qghnn,
  author = {Wenxuan Wang and Jinjing Shi and Nan Guan and Shichao Zhang and Xuelong Li},
  title = {{QGHNN: A Quantum Graph Hamiltonian Neural Network}},
  journal = {IEEE Transactions on Emerging Topics in Computational Intelligence},
  year = {2025},
  doi = {10.1109/tetci.2025.3604791},
}
Figure