Hyperbolic HAN (HHGAT) Framework

• The paper introduces HHGAT, a framework that uses hyperbolic geometry and Möbius algebra to model complex heterogeneous graphs.
• It employs the Poincaré ball model and automatic meta-path sampling to capture hierarchical structures and power-law distributions with minimal distortion.
• Empirical results demonstrate that HHGAT outperforms Euclidean and homogeneous baselines on node classification and clustering tasks across benchmarks like IMDB, DBLP, and ACM.

Hyperbolic Heterogeneous Graph Attention Networks (HHGAT) is a framework designed for representation learning in heterogeneous graphs by embedding node and meta-path information within a unified hyperbolic geometry, specifically leveraging the Poincaré ball model. Unlike prior approaches that constrain node embeddings to Euclidean spaces, HHGAT exploits the geometric properties of hyperbolic space, which is inherently suitable for complex structures commonly present in real-world heterogeneous graphs, such as hierarchical relationships and power-law degree distributions. By combining Möbius algebraic operations, meta-path instance sampling, and multi-level attention, HHGAT achieves state-of-the-art performance in node classification and clustering tasks across multiple datasets (Park et al., 2024).

1. Motivation for Hyperbolic Representations

Heterogeneous graphs—such as bibliographic and social networks—frequently exhibit hierarchical organization (e.g., research areas branching into sub-areas, then into papers, then authors) and power-law degree distributions, where a minority of nodes have disproportionately high connectivity. Embedding such topologies in Euclidean space leads to significant distortion due to the only linear expansion in Euclidean geometry. By contrast, spaces of constant negative curvature—hyperbolic spaces—expand exponentially with radius, efficiently accommodating scale-free and hierarchical patterns. Hyperbolic representation therefore enables faithful modeling of both node-level and complex meta-path instance relationships in heterogeneous graphs (Park et al., 2024).

2. The Poincaré Ball Model and Hyperbolic Operations

HHGAT operates in the $n$-dimensional Poincaré ball $\mathbb{D}^{n,c}$ of curvature $-c$ ($c > 0$), formally defined as: $$\mathbb{D}^{n,c} = \left\{ x \in \mathbb{R}^n : c\|x\|^2 < 1 \right\}, \quad g^c_x = (\lambda^c_x)^2 I_n,$$ where $\lambda^c_x = 2/(1 - c\|x\|^2)$. The core operations underlying HHGAT’s parameterizations are:

• Möbius addition:

$$x \oplus_c y = \frac{ (1 + 2c\,\langle x, y\rangle + c\|y\|^2) x + (1 - c\|x\|^2) y } { 1 + 2c\,\langle x, y\rangle + c^2\|x\|^2\|y\|^2 }$$

• Exponential map at the origin: $$\exp_o^c(v) = \tanh( \sqrt{c}\|v\| )\frac{v}{\sqrt{c}\|v\| }$$
• Logarithm map at the origin: $$\log_o^c(x) = \frac{1}{\sqrt{c}} \mathrm{artanh}( \sqrt{c}\|x\| )\frac{x}{\|x\| }$$
• Hyperbolic distance: $$d_c(x, y) = \frac{2}{\sqrt{c}} \mathrm{artanh}( \sqrt{c} \| (-x)\oplus_c y \| )$$
• Möbius scalar-matrix multiplication: $\mathbb{D}^{n,c}$0
• Hyperbolic activation: $\mathbb{D}^{n,c}$1

Operating entirely in hyperbolic geometry, these transformations permit attention and aggregation in non-Euclidean space without requiring auxiliary approximations or mappings.

3. HHGAT Architecture and Meta-path Embedding Workflow

The HHGAT framework consists of a multi-stage pipeline mapping raw node features to output embeddings:

3.1 Meta-path Instance Sampling:

Given target node $\mathbb{D}^{n,c}$2, a breadth-first search (BFS) of length $\mathbb{D}^{n,c}$3 is executed over the heterogenous schema (respecting link-type constraints), yielding the set $\mathbb{D}^{n,c}$4 of meta-path instances. No hand-crafted metapath design is needed.

3.2 Meta-path Instance Embedding:

• Euclidean Concatenation: For each meta-path instance $\mathbb{D}^{n,c}$5 (with $\mathbb{D}^{n,c}$6), concatenate node features: $\mathbb{D}^{n,c}$7.
• Exponential Map: Push to hyperbolic space: $\mathbb{D}^{n,c}$8.
• Hyperbolic Linear Transformation and Bias:

$\mathbb{D}^{n,c}$9.

3.3 Instance-level Attention & Aggregation:

• Compute hyperbolic attention for each instance $-c$0 in meta-path $-c$1:

$-c$2

• Aggregate via Möbius-weighted sum:

$-c$3

3.4 Inter-meta-path Attention and Fusion:

• For each meta-path type $-c$4, compute

$-c$5

• The final embedding is then:

$-c$6

4. Training and Optimization

HHGAT outputs a final embedding $-c$7 per node, which is mapped through a linear classifier for classification tasks: $-c$8 For supervised scenarios, the cross-entropy loss is applied to the labeled subset $-c$9: $c > 0$0 where $c > 0$1 is the one-hot ground truth label vector. All manifold-tangent conversions and trainable operations utilize the hyperbolic-exponential/logarithmic maps and Möbius algebraic structures, precluding the need for additional approximation techniques during optimization.

5. Empirical Evaluation and Comparative Analysis

HHGAT’s empirical performance was evaluated on the following heterogeneous benchmark datasets:

• IMDB (movies, directors, actors; 3 classes)
• DBLP (authors, papers, conferences; 4 classes)
• ACM (papers, authors, subjects; 3 classes)

The method was assessed against Euclidean homogeneous (GCN, GAT), hyperbolic homogeneous (HGCN), Euclidean heterogeneous (HAN, MAGNN, GTN, HetSANN, HGT, GraphMSE, Simple-HGN), and hyperbolic heterogeneous (SHAN) baselines.

Metrics used included Macro-F1, Micro-F1 for classification, and NMI, ARI for clustering. HHGAT consistently outperformed all baselines across both classification and clustering, with improvements typically in the 1–2 percentage points range on Macro/Micro-F1 and NMI/ARI (see Tables 3–4 of (Park et al., 2024)).

A curvature sensitivity analysis demonstrated optimal performance for negative curvatures near $c > 0$2 (IMDB), $c > 0$3 (DBLP), and $c > 0$4 (ACM). As $c > 0$5 (Euclidean limit), accuracy deteriorates, while excessive curvature also results in degraded performance, underscoring the critical role of matching curvature to the graph’s scale-free structure.

6. Technical Synopsis and Significance

HHGAT integrates meta-path sampling, hyperbolic embedding via Möbius operations, and two-level attention mechanisms in a mathematically consistent framework for heterogeneous graph representation. Its architecture is distinguished by:

• Automatic meta-path instance sampling without hand-crafted path design.
• Embedding both paths and node features into hyperbolic space using Möbius algebra.
• Attention and aggregation mechanisms formulated directly in hyperbolic geometry.
• End-to-end training under standard cross-entropy, fully leveraging the geometric representation power of negatively curved spaces.

This methodology facilitates accurate, low-distortion embeddings of complex, real-world heterogeneous graphs, yielding improved learning and inference performance in classification and clustering tasks where previous Euclidean or non-hyperbolic techniques exhibited measurable limitations (Park et al., 2024).