HSMGNN Hyperbolic S2 Rotation Group Multi-Channel Graph Neural Network for Cross-lingual Entity Alignment

Entity alignment—a crucial task in knowledge graph research—is often hampered by the inherent heterogeneity between the graphs being aligned. These differences pose serious challenges for existing alignment methods, highlighting the need for more innovative and adaptable solutions. To address this, we propose a novel framework that integrates with graph neural networks: the Hyperbolic S² Rotation Group in Multi-Channel Graph Neural Network (HSMGNN). HSMGNN is designed to effectively capture the diverse and complex structures of heterogeneous knowledge graphs, offering a streamlined and powerful approach to entity alignment. HSMGNN operates across multiple channels, each equipped with a distinct relation-weighting strategy for encoding knowledge graphs. One channel uses self-attention mechanisms to complete the knowledge graph by filling in missing data within each graph. Simultaneously, another channel applies cross-graph attention to pinpoint and filter out entities that are irrelevant to the alignment task. These diverse features are then aggregated using pooling techniques to form a unified, high-quality representation. What distinguishes HSMGNN is its ability to infer and transfer rule-based knowledge between graphs, ensuring coherent and consistent completion on both sides. This end-to-end framework not only boosts performance in entity alignment tasks but also bridges structural differences between heterogeneous graphs in a principled way. Extensive experiments on four benchmark datasets (DBP(zh-en), DBP(ja-en), DBP(fr-en), DBP(wd)) demonstrate that HSMGNN achieves significant performance gains over state-of-the-art methods. Specifically, on the DBP(zh-en) dataset, HSMGNN attains an H@1 of 0.533, an H@10 of 0.841, and an MRR of 0.641, which are 4.0% (H@1), 0.8% (H@10), and 4.9% (MRR) higher than the second-best model MuGNN. On the more challenging DBP(ja-en) and DBP(fr-en) datasets, HSMGNN shows even more remarkable improvements, with H@1 reaching 0.626, H@10 hitting 0.927, and MRR reaching 0.731—surpassing MuGNN by 25.0% (H@1), 8.2% (H@10), and 17.7% (MRR). For the DBP(wd) dataset, HSMGNN achieves an H@1 of 0.625, an H@10 of 0.912, and an MRR of 0.725, outperforming the previous leading model MuGNN by 1.5% (H@1), 1.7% (H@10), and 1.5% (MRR). In short, HSMGNN marks a meaningful step forward in entity alignment, offering a robust, efficient, and scalable solution for aligning knowledge graphs in complex real-world scenarios.