Synthetic Many-to-Many Data Generation With Private Graph Factorization

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for generating synthetic data sets with many-to-many relationships lack scalability and privacy preservation, often requiring users to compromise between neglecting these relationships or sacrificing privacy.

Innovation Solution

A method that generates synthetic data by first creating a synthetic graph based on associations between tables, then determining attributes using conditional models conditioned on the graph, ensuring differential privacy and efficiency through techniques from graph theory and representation learning.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If previous methods generate synthetic data with many-to-many relationships, then the synthetic data can represent real data relationships, but the generation process becomes computationally complex and non-scalable

Engineering Contradiction:
Improverepresentation of many-to-many relationshipsVSAvoidgeneration process complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent segments the synthetic data generation process into distinct stages: first generating the graph structure (edges and nodes), then conditionally generating attribute values based on the graph. This segmentation separates the complex relational structure generation from the attribute value generation, making the overall process more manageable and scalable while maintaining accurate representation of many-to-many relationships.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary action by generating the graph structure (edges and nodes) before generating the attribute values. The graph is generated first using the association data, and then attribute values are conditionally generated based on this pre-established graph structure. This preliminary generation of relational structure enables efficient subsequent attribute generation and improves overall computational efficiency.

Inventive Principle:
Principle #10Preliminary action

2Object-affected harmful factors

If previous methods prioritize privacy protection, then user privacy is maintained, but the synthetic data generation becomes slower and less efficient

Engineering Contradiction:
Improveprivacy protectionVSAvoidgeneration speed
Core Design Contradiction:
Object-affected harmful factorsVSProductivity

Solution Approach 1:

The patent applies preliminary action by generating the graph structure and privacy-preserving constraints before generating attribute values. The graph is constructed first using differentially private techniques, and then attribute values are generated conditioned on this pre-computed graph. This ordering enables efficient privacy protection during the structural generation phase, which is computationally less intensive than generating all attributes while maintaining privacy.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent segments the generation process into a privacy-critical phase (graph generation with differential privacy) and a faster attribute generation phase (conditional generation given the graph). By separating these phases, the patent can apply rigorous privacy protection where most needed during graph construction, while allowing faster, less restrictive generation during the attribute value phase, thereby improving overall productivity while maintaining privacy.

Inventive Principle:
Principle #1Segmentation

3Productivity

If previous methods generate synthetic data quickly, then productivity is improved, but scalability and applicability to diverse data sets are compromised

Engineering Contradiction:
Improvegeneration speedVSAvoidapplicability to different data sets
Core Design Contradiction:
ProductivityVSAdaptability or versatility

Solution Approach 1:

The patent achieves universality by creating a general framework that can handle diverse data sets with different many-to-many relationship structures. The graph-based approach with conditional generation models can be applied to various domains (healthcare, finance, social networks) without requiring domain-specific algorithms. The system generates synthetic data that preserves statistical properties and relationships across different data types, making it highly adaptable and scalable.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Solution Approach 2:

The patent utilizes parameter changes by adjusting the conditional generation models and graph generation parameters to adapt to different data sets. The system can modify the complexity of generation models, the level of privacy protection, and the graph generation parameters based on the specific characteristics of the input data, enabling both speed and scalability across diverse applications while maintaining accuracy and privacy.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4526793B1Synthetic generation of data with many to many relationships
Publication Date: 2026.03.11 SAS INSTITUTE INC
  • EP4526793B1 patent drawingFigure 1
  • EP4526793B1 patent drawingFigure 2
  • EP4526793B1 patent drawingFigure 3

AI summary

Embodiments described herein relate to the efficient generation of synthetic datasets that represent many-to-many relationships. In particular, certain embodiments implement a particular factorization for many-to-many generative models, which leads to a scalable generation framework by combining random graph theory and representation learning. Further embodiments we extend the framework to establish the notion of differential privacy within the synthetically generated data. The embodiments described herein are therefore able to generate synthetic datasets efficiently while preserving information within and across many-to-many datasets with improved accuracy.