Graph Neural Networks and their Extensions Based on Hyperbolic Geometry: a Review
P. B. Petrenko, А.С. Толпыгин · INFORMACIONNYE TEHNOLOGII · 2025
Graph neural networks generalize traditional neural networks for graph-structured data and have attracted wide attention due to their impressive capabilities. They are in demand in machine learning, as they can work with heterogeneous information, help to identify relationships between events and data, provide robustness to incomplete, unclear and noisy data, allow to analysis of large amounts of data and structures in the form of knowledge graphs. Training of graph neural networks in hyperbolic space has gained popularity in recent years due to their ability to model graphs with hidden hierarchical data, and because they are more compact and denser. The purpose of applying non-Euclidean geometry theory in modifying graph neural networks is to ensure their stability and better performance compared to Euclidean neural networks. The review of research and development is devoted to the analysis of modern achievements in the field of graph neural networks creation and consideration of accompanying problems. Particular attention is paid to the issues of graph embedding and their representation in non-Euclidean space. This is due to the fact that the performance of models is largely determined by the embedding algorithms and the choice of geometric space for representing graph structures. The prospectivity and efficiency of this approach have been confirmed by works on the creation of hyperbolic convolutional networks of graphs with the property of controlling the curvature of the space in each layer. At the same time, today there is a need for a more detailed consideration of the possibilities of graph neural networks based on the application of methods of hyperbolic geometry because of the insufficient attention paid to these issues in the domestic literature.