Graph Data Generation With Joint Node–Edge Attribute Diffusion

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing graph diffusion models fail to accurately model the interdependence of node attributes and edge attributes in graph data, leading to a lack of information and accuracy, especially in highly interdependent graph data.

Innovation Solution

A method is provided for training a machine learning diffusion model to generate graph data by performing a joint diffusion process that simultaneously solves forward-time and reverse-time stochastic differential equations for both node and edge attributes, using a score-approximation model that combines node and edge attributes through graph neural networks to capture their mutual dependence.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If separate diffusion processes are used for node attributes and edge attributes, then the model structure is simpler, but the interdependence between node and edge attributes is not accurately modeled

Engineering Contradiction:
Improvemodel structureVSAvoidinterdependence information
Core Design Contradiction:
Device complexityVSLoss of information

Solution Approach 1:

The patent merges the separate diffusion processes for node attributes and edge attributes into a unified joint diffusion process. The diffusion model simultaneously diffuses both node attributes and edge attributes through coupled stochastic differential equations, allowing the processes to interact and model their mutual dependence while maintaining a manageable overall structure.

Inventive Principle:
Principle #5Merging (Combining)

2Measurement precision

If a joint diffusion process is used to model interdependence, then information accuracy improves, but computational complexity increases

Engineering Contradiction:
Improveinformation accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the joint diffusion process into distinct but coupled components: node attribute diffusion and edge attribute diffusion. Each component has its own stochastic differential equation with clearly defined terms, allowing for modular implementation and computation while preserving the interdependence through coupling mechanisms.

Inventive Principle:
Principle #1Segmentation

3Productivity

If existing graph diffusion models are used, then generation speed is maintained, but accuracy in highly interdependent graph data deteriorates

Engineering Contradiction:
Improvegeneration speedVSAvoidgeneration accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent incorporates feedback mechanisms where the diffusion of node attributes and edge attributes mutually influence each other through the coupled stochastic differential equations. The score functions for both node and edge attributes depend on their respective current states, creating a feedback loop that continuously refines the generation accuracy for highly interdependent graph data while maintaining efficient sampling.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP4607408A1Using diffusion model to generate graph data
Publication Date: 2025.08.27 ROBERT BOSCH GMBH
  • EP4607408A1 patent drawingFigure 1
  • EP4607408A1 patent drawingFigure 2
  • EP4607408A1 patent drawingFigure 3

AI summary

Methods are provided for training and using a machine learning diffusion model to generate graph data based on samples from a data distribution as input. The diffusion model comprises one or more diffusion layers, and the graph data comprises node attributes and edge attributes. The training method comprises a diffusion process, comprising a forward- and a reverse-time pass, to learn parameters of the diffusion layers, and a joint diffusion process, comprising solving a forward- and a reverse-time stochastic differential equation. Both equations are based on both the node and edge attributes, the reverse-time equation being additionally based on the learnt parameters of the diffusion layers. Both equations are solved for both for the node and the edge attributes simultaneously. The trained diffusion model is provided for use, in which use method the trained diffusion model repeatedly performs the reverse-time pass to obtain graph data based on input samples.