Reviving Life on the Edge: Joint Score-Based Graph Generation of Rich Edge Attributes

TL;DR

The paper tackles the challenge of generating graphs with rich edge attributes by proposing a joint score-based diffusion framework that evolves node and edge features together in a unified SDE for graphs $G=(X,E)$. A specialized Graph Neural Network combines a Graph Neural Module and multi-head Attention to propagate information across nodes, edges, and adjacency, enabling mutual dependence during diffusion. Key contributions include (i) a joint score model for nodes and edges, (ii) an edge-aware architectural bias that effectively propagates edge information, and (iii) extensive evaluation on edge-critical benchmarks (synthetic MDP grids and nuScenes traffic scenes) with ablations showing the importance of joint diffusion and edge-centric components. The approach yields superior edge-related metrics and provides a foundation for more expressive graph generative models with practical impact on traffic scene generation and related edge-dominated tasks.

Abstract

Graph generation is integral to various engineering and scientific disciplines. Nevertheless, existing methodologies tend to overlook the generation of edge attributes. However, we identify critical applications where edge attributes are essential, making prior methods potentially unsuitable in such contexts. Moreover, while trivial adaptations are available, empirical investigations reveal their limited efficacy as they do not properly model the interplay among graph components. To address this, we propose a joint score-based model of nodes and edges for graph generation that considers all graph components. Our approach offers three key novelties: \textbf{(1)} node and edge attributes are combined in an attention module that generates samples based on the two ingredients, \textbf{(2)} node, edge and adjacency information are mutually dependent during the graph diffusion process, and \textbf{(3)} the framework enables the generation of graphs with rich attributes along the edges, providing a more expressive formulation for generative tasks than existing works. We evaluate our method on challenging benchmarks involving real-world and synthetic datasets in which edge features are crucial. Additionally, we introduce a new synthetic dataset that incorporates edge values. Furthermore, we propose a novel application that greatly benefits from the method due to its nature: the generation of traffic scenes represented as graphs. Our method outperforms other graph generation methods, demonstrating a significant advantage in edge-related measures.

Figures (12)

• Figure 1: Simulation of a possible traffic scene with two cars, a truck, and a motorcycle.
• Figure 2: The corresponding graph representation of the scene.
• Figure 4: Our architecture. On the left, we show the score modules (Eq. \ref{['eq:scores']}). On the right (Blue), our GCN module is constructed of GNM (Eq. \ref{['eq:GNM']}) layers. On the right (Red), the GMH module is constructed of attention (ATTN) layers.
• Figure 5: Our ablation study shows that GDSS-E and the other variants yield inferior distributions compared to our approach concerning the ground-truth data distribution estimation.
• Figure 6: A qualitative comparison between the original data (left), GDSS-E (middle), and our method (right) on the deterministic MDP grid maze dataset. Blocks are colored in blue (squares), and start and finish nodes in yellow (diamonds) and green (triangles), respectively. Our graphs consistently have four blocks, one start node, and one finish node, as required.