Abstract:Diffusion models have made significant contributions to computer vision, sparking a growing interest in the community recently regarding the application of them to graph generation. Existing discrete graph diffusion models exhibit heightened computational complexity and diminished training efficiency. A preferable and natural way is to directly diffuse the graph within the latent space. However, due to the non-Euclidean structure of graphs is not isotropic in the latent space, the existing latent diffusion models effectively make it difficult to capture and preserve the topological information of graphs. To address the above challenges, we propose a novel geometrically latent diffusion framework HypDiff. Specifically, we first establish a geometrically latent space with interpretability measures based on hyperbolic geometry, to define anisotropic latent diffusion processes for graphs. Then, we propose a geometrically latent diffusion process that is constrained by both radial and angular geometric properties, thereby ensuring the preservation of the original topological properties in the generative graphs. Extensive experimental results demonstrate the superior effectiveness of HypDiff for graph generation with various topologies.

What problem does this paper attempt to address?

This paper attempts to solve two main problems in graph generation tasks: 1. **High computational complexity**: Existing discrete graph diffusion models usually have high time and space complexity due to the discreteness, sparsity and other topological characteristics of non - Euclidean structures. In particular, when dealing with fully - connected noisy graphs, these models will lose structural information and underlying topological properties. 2. **Anisotropy of non - Euclidean structures**: The non - Euclidean embeddings of graph data are anisotropic in the continuous latent space, which means that the traditional isotropic diffusion process cannot effectively capture and preserve the topological information of the graph. Specifically, node embeddings show significant anisotropy in Euclidean space, which will lead to the loss of useful structural information during the denoising process. To solve these problems, the authors propose a new geometric latent diffusion framework - **HypDiff** (Hyperbolic Geometric Latent Diffusion Model). The main contributions of this model include: - **Introducing hyperbolic geometric latent space**: Establish an interpretable geometric latent space based on hyperbolic geometry and define an anisotropic latent diffusion process applicable to graphs. Hyperbolic geometry can better represent tree - like or hierarchical structures and tends to be isotropic globally while retaining anisotropy locally. - **Designing a new geometric latent diffusion process**: Design the diffusion process through radial and angular geometric constraints to ensure that the original topological properties of the generated graph are preserved. Specifically, use radial measures for approximate diffusion and guide the diffusion model to capture the details of the graph structure more finely through angular constraints. Through the above methods, HypDiff can effectively deal with the anisotropy problem of non - Euclidean structures in graph generation tasks and significantly improve the quality and performance of the generated graphs. Experimental results show that HypDiff performs better than existing methods on multiple synthetic and real - world datasets. ### Formula summary 1. **Hyperbolic auto - encoding optimization formula**: \[ L_{\text{HAE}} = -E_{q_\phi(z_x|x)} \left[ \log p_\xi(x|z_x) \right] \] where $\log_c$ is the logarithmic mapping from the North Pole (origin) to the tangent space in hyperbolic space. 2. **Geometric diffusion process formula**: \[ x_t = \sqrt{\alpha_t} x_0 + \sqrt{1 - \alpha_t} z + \delta \tanh\left(\sqrt{c} \lambda_c \frac{t}{T_0}\right) x_0 \] where $\epsilon$ is Gaussian noise, $\delta$ is the radial epidemic coefficient that controls the diffusion intensity of each node in hyperbolic space, and $T_0$ is a constant that controls the radial growth rate. 3. **Denoising network loss function**: \[ L_{\text{HDM}} = E \left[ \| f_\theta(X_t, A, t) - X_0 \|^2 \right] \] Through the application of these formulas, the HypDiff model can efficiently deal with the challenges in graph generation tasks and show superior performance in experiments.

Hyperbolic Geometric Latent Diffusion Model for Graph Generation

Hyperbolic Graph Diffusion Model

Graphusion: Latent Diffusion for Graph Generation

Discrete-state Continuous-time Diffusion for Graph Generation

Generative Diffusion Models on Graphs: Methods and Applications

Autoregressive Diffusion Model for Graph Generation

Continuous Geometry-Aware Graph Diffusion via Hyperbolic Neural PDE

Topology-Aware Latent Diffusion for 3D Shape Generation

GLAD: Improving Latent Graph Generative Modeling with Simple Quantization

Graph Generation with Diffusion Mixture

GraphGDP: Generative Diffusion Processes for Permutation Invariant Graph Generation

Diffusion-based Graph Generative Methods

Sparse Training of Discrete Diffusion Models for Graph Generation

Fast Graph Generation via Spectral Diffusion

Directional diffusion models for graph representation learning

HYGENE: A Diffusion-based Hypergraph Generation Method

Advancing Graph Generation through Beta Diffusion

H-Diffu: Hyperbolic Representations for Information Diffusion Prediction

Latent Diffusion Models for Structural Component Design

Geometric Latent Diffusion Models for 3D Molecule Generation

A Geometry Aware Diffusion Model for 3D Point Cloud Generation