Graph Geometry Interaction Learning

1 Introduction

• Real-world graphs can have scale-free or hierarchical structure, for which Euclidean embeddings may incur large distortion while hyperbolic spaces offer a low-dimensional alternative.The introduction also describes graphs with regular structures that are better suited to Euclidean space and hierarchical structures that are better suited to hyperbolic space.Source: §1 · p. 1, Figure 1 · p. 2
• Prior approaches assume shared curvature or learn separate Euclidean and hyperbolic embeddings, which may be distorted by nodes with different structural properties.The paper argues that single-space embeddings can be suboptimal for graphs whose nodes exhibit different structures.Source: §1 · p. 2
• GIL jointly embeds graphs in Euclidean and hyperbolic spaces through interaction learning while maintaining each space’s conformal invariance.Its stated aim is to use Euclidean operational convenience while preserving hyperbolic modeling of complex structure.Source: §1 · p. 2
• GIL uses exponential and logarithmic maps for cross-space propagation, distance-aware hyperbolic attention, adaptive feature interaction, and node-customized probability assembling.The assembling weights determine the importance of each geometry’s embedding for downstream tasks.Source: §1 · p. 2
• GIL achieves state-of-the-art performance on five benchmarks for both node classification and link prediction.The introduction also reports extensive experiments verifying its effectiveness on both tasks.Source: §1 · p. 2
• The paper claims, to the authors’ best knowledge, to be the first to jointly exploit Euclidean and hyperbolic graph embeddings for complex graph structure.Source: §1 · p. 2

2 Notation and Preliminaries

• Hyperbolic space is defined as a constant-negative-curvature Riemannian manifold, with a tangent space providing a first-order local approximation around each point.Source: §2 · p. 3
• The paper adopts the Poincaré ball model and uses exponential and logarithmic maps to connect hyperbolic space with tangent space.The passage describes the model as compact and notes that it has principled generalizations of basic operations and closed-form expressions for basic objects such as distance.Source: §2 · p. 3
• Graph attention networks perform attentional message propagation and node updating, with attention coefficients parameterized by a trainable weight vector.The passage identifies one-hop neighbors and a trainable transformation in the update setup.Source: §2 · p. 3

3 Graph Geometry Interaction Learning

• GIL integrates geometry-specific features and category probabilities to produce final predictions.Its framework combines feature interaction across spaces with probability assembling.Source: §3 · p. 3

3.1 Geometry Feature Interaction

• Neighbor information propagates in both spaces; hyperbolic messages use distance-guided attention on the tangent space.Hyperbolic features are mapped into tangent space for distance-guided self-attention and then mapped back; the method also describes tangent-space aggregation as supporting permutation invariance.Source: Figure 2 · p. 4, §3.1 · pp. 4–5
• Node features are adaptively adjusted using similarity between geometric embeddings, while fused embeddings remain in their original spaces and preserve their conformal structure.The method measures similarity using hyperbolic and Euclidean distances.Source: §3.1 · p. 5

3.2 Hyperplane-Based Logistic Regression and Probability Assembling

• Hyperplane-based classification and node-level probability weights combine Euclidean and hyperbolic predictions; link prediction uses the same assembling scheme for edge probabilities.For link prediction, the probabilities are assembled using the difference between endpoint embeddings.Source: §3.2 · pp. 5–6

3.3 Model Analysis

• The end-to-end model uses cross-entropy training, with negative sampling for link prediction, and message propagation has time complexity O(NFF' + |E|F').The distance-aware attention operation can be parallelized across edges, while remaining computations can be parallelized across nodes.Source: §3.3 · pp. 6–7

4.1 Experiments Setup

• The five benchmarks span different hyperbolicities, with Disease the exception: it is a tree with 0-hyperbolic structure.The datasets are Disease, Airport, Cora, Pubmed, and Citeseer; Table 1 summarizes their sizes, features, classes, and hyperbolicity distributions.Source: §4.1 · p. 7, Table 1 · p. 7
• The evaluation compares GIL with shallow, neural-network, Euclidean GNN, and hyperbolic GNN baselines.The setup also includes EucHyp, a simple integration of Euclidean and hyperbolic embeddings learned from GAT and HGAT.Source: §4.1 · pp. 7–8

4.2 Experimental Results

• GIL achieves the best performance on all five datasets in both node classification and link prediction, outperforming the simple EucHyp combination.Table 2 reports link prediction in ROC AUC and node classification in accuracy; the prose specifies F1 for Disease and Airport.Source: Table 2 · p. 7, §4.2 · p. 8
• Feature-fusion variants outperform no fusion, and dual feature fusion gives the best node-classification performance.Without distance attention, classification accuracy decreases across datasets and variance increases.Source: Table 3 · p. 8, Figure 4 · p. 8, §4.2 · p. 9
• On Disease, hyperbolic weights are greater overall; on Citeseer, both geometries contribute, with boundary nodes generally receiving more hyperbolic weight.Figure 3 encodes greater probability weight with darker node color.Source: Figure 3 · p. 8, §4.2 · p. 9

5 Conclusion

• GIL combines hyperbolic and Euclidean topological features for graph representation.The method uses distance-aware propagation and interaction learning, and adaptively weights the geometric embeddings for each node.Source: §5 · p. 9
• GIL allocates different importance weights to geometric embeddings for each node through an adaptive scheme.The conclusion identifies this alongside distance-aware propagation as a core component of the method.Source: §5 · p. 9
• GIL achieves state-of-the-art performance across five benchmarks on node classification and link prediction.The conclusion also reports theoretical derivation and experimental analysis validating the method's effectiveness.Source: §5 · p. 9

• The model is intended to capture different intrinsic features across Euclidean and hyperbolic spaces for a wider range of complex graphs.Source: Impact Satement · p. 9
• Potential applications include social influence prediction, traffic flow prediction, and drug discovery.Source: Impact Satement · p. 9
• The model, like many GNNs, provides limited explanation for its predictions.The authors advocate further interpretability work to support GNN use in more critical medical applications.Source: Impact Satement · p. 9