New Method Accelerates Heterogeneous Graph Neural Network Training

Training Heterogeneous Graph Neural Networks (HGNNs) on massive, complex datasets is often computationally intensive, limiting their practical application despite their power in modeling diverse systems. Current graph condensation techniques, primarily designed for simpler homogeneous graphs, are too slow for heterogeneous settings due to their reliance on computationally demanding gradient matching or bilevel optimization. Researchers have developed HGC-RC, a new framework to address these limitations. HGC-RC efficiently condenses large heterogeneous graphs by first extracting semantically rich node embeddings using a lightweight propagation method. It then employs a "role-aware" hybrid clustering strategy. This strategy involves class-partitioned clustering for labeled nodes to maintain class distributions and unsupervised type-wise clustering for unlabeled nodes to preserve crucial cross-type connections. Finally, a compact, representative graph is reconstructed from these clusters. Extensive experiments show HGC-RC significantly improves learning efficiency and outperforms state-of-the-art baselines, making HGNN training more feasible for large-scale applications.