Geometry-aware Distance Measure for Diverse Hierarchical Structures in Hyperbolic Spaces
Pengxiang Li, Yuwei Wu, Zhi Gao, Xiaomeng Fan, Wei Wu, Zhipeng Lu, Yunde Jia, and Mehrtash Harandi · 2025
Read Geometry-aware Distance Measure for Diverse Hierarchical Structures in Hyperbolic Spaces by Pengxiang Li, Yuwei Wu, Zhi Gao, Xiaomeng Fan, Wei Wu, Zhipeng Lu, Yunde Jia, and Mehrtash Harandi (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-18533.txt.
Introduction
Pengxiang Li and co-authors rework distance in curved learning spaces in a 2025 paper. Hyperbolic spaces suit tree-like data, so learners use them for ranked sets. Most set one fixed distance rule for all points, as if every tree had the same shape.
Real trees vary, so the team let the distance rule adapt to each pair of points. Each pair gets its own bend and map into a fitting space. Tricks with low-rank math and hard-pair picks keep the cost in check, with a proven error bound. Tests on image sorting, ranked sorting, and few-shot tasks beat fixed-rule rivals across the board.
Catalog record
| CREATOR | Pengxiang Li, Yuwei Wu, Zhi Gao, Xiaomeng Fan, Wei Wu, Zhipeng Lu, Yunde Jia, and Mehrtash Harandi |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 63 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2506-18533
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Geometry-aware Distance Measure for Diverse Hierarchical Structures in Hyperbolic Spaces Pengxiang Li, Yuwei Wu, Zhi Gao, Xiaomeng Fan, Wei Wu, Zhipeng Lu, Yunde Jia, and Mehrtash Harandi Abstract Learning in hyperbolic spaces has attracted increasing attention due to its superior ability to model hierarchical structures of data. Most existing hyperbolic learning methods use fixed distance measures for all data, assuming a uniform hierarchy across all data points. However, real-world hierarchical structures exhibit significant diversity, making this assumption overly restrictive. In this paper, we propose a geometry-aware distance measure in hyperbolic spaces, which dynamically adapts to varying hierarchical structures. Our approach derives the distance measure by generating tailored projections and curvatures for each pair of data points, effectively mapping them to an appropriate hyperbolic space. We introduce a revised low-rank decomposition scheme and a hard-pair mining mechanism to mitigate the computational cost of pair-wise distance computation without compromising accuracy. We present an upper bound on the low-rank approximation error using Talagrand’s concentration inequality, ensuring theoretical robustness. Extensive experiments on standard image classification (MNIST, CIFAR-10/100), hierarchical classification (5-level CIFAR-100), and few-shot learning tasks (mini-ImageNet, tiered-ImageNet) demonstrate the effectiveness of our method. Our approach consistently outperforms learning methods that use fixed distance measures, with notable improvements on few-shot learning tasks, where it achieves over 5% gains on mini-ImageNet. The results reveal that adaptive distance measures better capture diverse hierarchical structures, with visualization showing clearer class boundaries and improved prototype separation in hyperbolic spaces. keywords Hyperbolic geometry, Metric learning, Few-shot learning, Image classification 1 Introduction Hyperbolic space is defined as a smooth Riemannian manifold with constant negative curvature. Compared with the commonly used Euclidean space, a notable property of hyperbolic spaces lies in the exponential growth of a ball’s volume relative to its radius, mirroring the exponential increase in the volume of hierarchical data with depth. Such a property enables the hyperbolic space to serve as a continuous analogue to trees ( Sala et al. 2018 ; Balazevic et al. 2019 ) , allowing for modeling hierarchical data with minimal distortion ( Sarkar 2011 ) . The use of hyperbolic spaces for data embeddings has shown superiority over Euclidean space across many applications such as classification ( Gao et al. 2021 ; Zhang et al. 2022 ; Gao et al. 2022 ; Atigh et al. 2025 ) , clustering ( Lin et al. 2022 ; Lin et al. 2023a ; Liu et al. 2025 ) , retrieval ( Ermolov et al. 2022a ) , detection ( Hong et al. 2024 ; Li et al. 2024 ; Gonzalez-Jimenez et al. 2025 ) , segmentation ( Hsu et al. 2021 ; Atigh et al. 2022a ) , multi-modal learning ( Hong et al. 2023 ; Long and van Noord 2023 ; Mandica et al. 2025 ) , and 3D vision ( Montanaro et al. 2022 ; Leng et al. 2023 ; Lin et al. 2023b ; Li et al. 2025 ) . Most existing hyperbolic learning methods employ a fixed distance measure, i.e. , the geodesic distance ( Khrulkov et al. 2020 ; Ermolov et al. 2022a ; Lin et al. 2023b ) , to assess the similarities between data points. Employing a fixed distance measure implicitly assumes a uniform hierarchical structure across all data points ( Behrstock et al. 2019 ) . However, this assumption does not always hold in real-world scenarios, as the hierarchical structures among data points are diverse and complex ( Pal et al. 2013 ) . Thus, using a fixed distance measure in real-world scenarios may cause data distortion for diverse hierarchical structures, leading to suboptimal performance. Figure 1: Modeling data with diverse hierarchical structures will benefit from adaptive distance measures. (a) The “dog-wolf” data pair has a simple semantic hierarchical structure, while the “aircraft carrier-school bus” data pair has a more complex structure. (b) A fixed distance measure fails to fit the hierarchical structures of the two data pairs simultaneously. (c) Flatter distance measures are suitable for the simple structure. (d) Steeper distance measures are suitable for the complex structure. The blue dashed lines represent distance measures in hyperbolic space. The solid lines represent the semantic subtrees. For example, as shown in Figure 1 (a), the hierarchical structures of data pairs can be reflected via connecting paths between two nodes on the corresponding graph/tree ( Behrstock et al. 2019 ) . The “dog-wolf” data pair has a simple semantic hierarchical structure, while the “aircraft carrier–school bus” data pair has a more complex structure. In this case, a fixed distance measure fails to accurately represent the two different hierarchical relationships, that is, the computed distances do not pass through the ancestors of the data pairs, leading to a misalignment between distances and hierarchical structures. As detailed in Figure 1 (b), the “dog-wolf” pair and “aircraft carrier–school bus” pair have different similarities with different hierarchical structures, but exhibit similar distance values in the embedding space if a fixed distance measure is used. Thus, employing adaptive distance measures for data with diverse hierarchical structures seems to be a natural choice, which could take the complexity of the hierarchical structures into consideration and dynamically fit the diverse hierarchical structures. As shown in Figure 1 (c) and (d), the adaptive distance measure produces shorter paths for the “dog-wolf” pair and longer paths for the “aircraft carrier–school bus” pair, better reflecting their respective hierarchical complexities. In this paper, we propose a geometry-aware distance measure that dynamically adapts to diverse hierarchical structures in hyperbolic spaces. The main idea is to learn to generate adaptive projections and curvatures for each pair of samples in hyperbolic spaces, thereby conforming to the hierarchical relationships between data points. To achieve this, we design a curvature generator to produce adaptive curvatures for different data pairs, and a projection matrix generator to map data pairs from the original hyperbolic space into adaptive hyperbolic spaces characterized by the newly produced curvature. By applying adaptive projections and curvatures to different sample pairs, we obtain our distance measures. Two challenges need to be solved in learning to generate geometry-aware distance measures in hyperbolic space. (1) Given that unique projections and curvatures are required for every data pair, reducing the computational cost becomes a critical consideration. (2) As a variety of hierarchical structures are encountered during training, the training stability for pair-wise learning is not easily maintained. To address the first challenge, we introduce a low-rank decomposition scheme and a hard-pair mining mechanism. The former reduces the computational complexity of the projection matrix via low-rank approximation. The latter eliminates easy samples by only generating geometry-aware distances for the remaining challenging data pairs. We theoretically prove that the approximation errors of low-rank decomposition are bounded with high probability. For the second challenge, we show that incorporating residual connections into the projection matrix during the generation process can effectively maintain training stability. Experimental results across standard image classification (MNIST, CIFAR-10/100), hierarchical classification (5-level CIFAR-100), and few-shot learning tasks (mini-ImageNet, tiered-ImageNet) confirm the effectiveness of our method in refining hyperbolic learning through geometry-aware distance measures. The experiments reveal that adaptive distance measures effectively capture diverse hierarchical structures, as evidenced by improved class separation in visualizations and better alignment between learned representations and underlying data hierarchies. The primary contributions of our