Hyperbolic Heterogeneous Graph Attention Networks
Jongmin Park, Seunghoon Han, Soohwan Jeong, Sungsu Lim
TL;DR
This work tackles the challenge of representing heterogeneous graphs with complex hierarchical and power-law structures in Euclidean space. It introduces Hyperbolic Heterogeneous Graph Attention Networks (HHGAT), which automatically sample metapath instances and learn their representations in hyperbolic space using hyperbolic attention mechanisms, followed by inter-metapath fusion. The model combines Euclidean-to-hyperbolic feature mapping, hyperbolic linear transforms, and curvature-aware attention to produce expressive node embeddings, achieving state-of-the-art performance on node classification and clustering across three real-world datasets and showing curvature as a tunable, dataset-dependent parameter. The results highlight the practical value of learning in hyperbolic space for heterogeneous graphs and point to future work on curvature regularization to better capture power-law distributions.
Abstract
Most previous heterogeneous graph embedding models represent elements in a heterogeneous graph as vector representations in a low-dimensional Euclidean space. However, because heterogeneous graphs inherently possess complex structures, such as hierarchical or power-law structures, distortions can occur when representing them in Euclidean space. To overcome this limitation, we propose Hyperbolic Heterogeneous Graph Attention Networks (HHGAT) that learn vector representations in hyperbolic spaces with meta-path instances. We conducted experiments on three real-world heterogeneous graph datasets, demonstrating that HHGAT outperforms state-of-the-art heterogeneous graph embedding models in node classification and clustering tasks.