Abstract:Heterogeneous graph neural networks have made significant progress, yet challenges such as excessive parameter size and high computational overhead remain. Quantum graph neural network (QGNN), by leveraging the advantages of quantum computing, offers a potential solution to enhance the computational efficiency of traditional graph neural networks. However, the complex semantic relations and diverse feature types inherent in heterogeneous graphs hinder the direct application of existing QGNNs, which are primarily designed for homogeneous graphs. To address this limitation, a quantum heterogeneous graph attention network (QHAN) is proposed. First, a dynamic graph decomposition strategy is introduced, wherein subgraphs are further partitioned based on the maximum number of available qubits on quantum hardware. A bidirectional amplitude encoding method is designed to encode quantum states. Building on this, a parameterized quantum circuit based on linear combinations of unitaries is constructed to compute multi-level semantic attention coefficients. This design enables parallel computation of node-level and semantic-level attention, while also supporting parameter sharing within the same level. Experiments on ten benchmark datasets demonstrate that the proposed QHAN significantly outperforms existing QGNN models in classification tasks. In addition, comparable performance to classical heterogeneous graph neural networks is achieved with fewer parameters.