Graph deep learning is rapidly transforming fields such as data mining and machine learning, but current methods often fail to fully account for the non-Euclidean structure inherent in graph data. Li Sun and Qiqi Wan of Beihang University and Suyang Zhou of East China Normal University, in collaboration with Zhenhao Huang and Philip S. Yu of North China Electric Power University, demonstrated that Riemannian geometry provides a powerful, principled framework for advancing the learning of graph representations. Their work established Riemann graph learning not as a set of isolated techniques, but as a unified paradigm that can address the limitations of existing approaches that often rely on external manifold formulations and restricted manifold types. By proposing an intrinsic manifold structure within graph neural networks, this study identifies critical gaps, proposes a structured research agenda, and ultimately provides a consistent perspective that inspires further exploration of Riemannian geometry as a foundational element for future graph learning research.
Scientists are fundamentally reimagining graph deep learning through the application of Riemannian geometry, a branch of mathematics concerned with curved spaces. Unlike traditional approaches that treat graphs as existing in a flat Euclidean space, this study establishes that modeling graphs on Riemannian manifolds, a generalization of Euclidean spaces through curved geometry, provides a more principled and expressive learning foundation. This is especially important given the inherent non-Euclidean structure of graphs. In graphs, the relationships between objects are complex and cannot be easily captured using standard deep learning techniques. This study reveals that there are significant gaps as current methodologies often focus on limited types of manifolds, especially hyperbolic spaces, and rely on external formulations that inherently embed graphs in high-dimensional Euclidean spaces. Instead, the core mission should be to equip graph neural networks with an inherent manifold structure to directly model the geometry of the graph within the curved space itself. This perspective requires a structured research question that includes manifold types, neural network architectures, and learning paradigms. This study provides a comprehensive classification of existing techniques, categorized by manifold type, such as hyperbolic spaces, spherical spaces, pseudo-Riemannian spaces, neural architectures, and learning paradigms. The resulting framework not only clarifies the landscape of Riemann graph learning but also highlights important areas for future investigation. Beyond a theoretical advance, this work suggests potential applications in a variety of fields, from recommender systems and social media analysis to molecular biology and physical interaction systems, promising more accurate and insightful data analysis. Graphs are ubiquitous in various data domains, and their non-Euclidean structure and complex interactions pose unique challenges to machine learning. This research establishes Riemannian geometry as a fundamental framework for learning graph representations and goes beyond viewing Riemannian geometry as a mere collection of techniques. This work reveals the critical need to provide graph neural networks with an intrinsic manifold structure. This is a largely unexplored area despite the recent integration of graph learning and Riemannian geometry. Eight representative manifolds are considered, including hyperbolas, spheres, constant curvatures, products, quotients, pseudo-Riemannian manifolds, Grassmann manifolds, and general manifolds, demonstrating the breadth of potential geometric approaches. Six neural architectures, graph convolutional networks, graph variational autoencoders, transformers, graph ODEs, denoising diffusion and SDE, and flow matching, are reviewed within this diverse context, highlighting how each can be adapted to exploit Riemannian geometry. In this study, we propose a conceptual framework that organizes Riemann graph learning along three dimensions: manifold type, neural architecture, and learning paradigm, providing a structured lens for understanding current approaches and identifying their limitations. This framework is not a comprehensive survey, but rather focuses on unresolved questions and new directions. This study highlights that while previous work has overemphasized hyperbolic spaces that are optimal for hierarchical data, real-world graphs are much more complex. Moreover, existing studies often generalize Euclidean formulations to manifolds without fully exploiting the unique concepts and tools of Riemannian geometry. The authors advocate a move towards intrinsic manifold properties in neural network design. This is a challenging but potentially powerful approach. This perspective is particularly relevant to the development of scientific artificial intelligence and graph-based models, where geometric priors fundamentally benefit scientific reasoning and reasoning. Riemannian geometry underpins this research approach to learning graph representations. Rather than treating graphs as simple collections of nodes and edges, this study positions graphs as essentially geometric objects that exist on a Riemannian manifold, a generalized shape that locally resembles Euclidean space but exhibits curvature globally. To investigate this, this study systematically catalogs the types of manifolds used in conjunction with graph neural networks, identifying hyperbolic manifolds, spherical manifolds, constant curvature spaces, product and quotient spaces, and pseudo-Riemannian manifolds as key areas of investigation. This methodology goes beyond simply applying Riemannian geometry. We focus on embedding graph neural networks within these intrinsic manifold structures. This involved examining existing models in detail and classifying them according to the type of manifold used and the specific neural architecture used to exploit its properties. For example, models that utilize hyperbolic geometry such as HGNN, p-VAE, and ROTE were analyzed for their implementation of hyperbolic distance metrics and their impact on representation learning. This detailed classification provides a nuanced understanding of how different geometric assumptions affect model performance. An important aspect of this work is its focus on eigenmanifold formulations. External approaches treat the manifold as embedded within a high-dimensional Euclidean space, which can result in the loss of valuable geometric information. Instead, this study favors direct manipulation of the manifold itself, utilizing the tools of differential geometry to define distances, angles, and other geometric properties. This is achieved through careful consideration of neural network layers that are designed to preserve geometric relationships and capture the inherent curvature of the graph structure. Scientists are increasingly recognizing that graphs are much more than collections of nodes and edges, with unique geometric structures that are much more noteworthy. For many years, machine learning has treated these networks as flat, abstract spaces, which is a simplification that limits its ability to accurately model complex relationships. This study convincingly argues for a shift to Riemannian geometry as the basic principle for learning graph representations. The challenges are not just technical. It’s conceptual, and we need to move away from thinking of graphs as tables of connections, and toward understanding the inherent dimensionality of graphs. The current focus on hyperbolic spaces, while useful, represents a potentially limiting special case. In parallel with innovative neural network architectures designed to exploit these shapes, it is important to investigate more broadly the types of manifolds that are the underlying shapes of these graphs. Existing methods often “force” the graph into a predefined manifold rather than allowing the geometry to emerge naturally from the data itself. The authors argue that this substantive approach has real benefits. Before Riemann graph learning can have a real impact on fields such as recommendation systems and drug discovery, it is essential to establish a robust theoretical foundation and develop scalable algorithms. The next wave of research may combine these geometric insights with new techniques such as diffusion models, unlocking entirely new capabilities in generative modeling and anomaly detection on complex networks. The hope is that the end result will not only be improved algorithms, but also a deeper understanding of the data itself.
