Graph neural networks (GNNs) excel at node classification tasks, but often assume homogeneity—that connected nodes share similar labels. This assumption does not hold true in many real-world heterogeneous graphs. Existing models for heterogeneous graphs primarily rely on pairwise relationships, overlooking multi-scale information available from higher-order structures. This leads to suboptimal performance, especially under noise caused by conflicting class information between nodes. To address this issue, we propose HPGNN, a novel model that integrates high-order personalized PageRank (PPR) with graph neural networks. HPGNN introduces an efficient high-order approximation of PPR to capture long-range, multi-scale node interactions. This approach reduces computational complexity and mitigates noise caused by surrounding information. By embedding high-order structural information into convolutional networks, HPGNN effectively models key interactions across multiple graph dimensions. Extensive experiments on benchmark datasets demonstrate the effectiveness of HPGNN. This model outperforms five of seven state-of-the-art methods on heterogeneous graphs in downstream tasks, while maintaining competitive performance on homogeneous graphs. HPGNN's ability to balance multi-scale information and maintain robustness against noise makes it a versatile solution for real-world graph learning problems.
