Many-to-Many Synthetic Data Generation via Private Graph Factorization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating synthetic data from datasets with many-to-many relationships lack scalability and privacy preservation, often requiring users to choose between neglecting these relationships or sacrificing privacy.
Innovation Solution
A method that generates synthetic graphs before determining attributes, using random graph theory and representation learning, ensuring differential privacy by injecting noise into the bipartite joint degree distribution, and employing conditional models to determine node attributes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If previous methods generate synthetic data from datasets with many-to-many relationships, then data fidelity is improved, but computational complexity increases and scalability deteriorates
Solution Approach 1:
The patent applies preliminary action by first generating the synthetic graph structure (edges and nodes) before generating the attribute values. This reverses the conventional approach where attributes are generated before relationships. By establishing the relational structure first, the computational complexity is reduced because subsequent attribute generation can be done independently for each node based on its position in the graph, rather than requiring complex joint modeling of all attributes and relationships simultaneously.
Solution Approach 2:
The patent segments the synthetic data generation process into distinct stages: (1) generating the synthetic graph structure from the bipartite joint degree distribution, (2) generating node embeddings from the graph structure, and (3) generating attribute values conditioned on node embeddings and graph structure. This segmentation allows each stage to be optimized independently and reduces overall computational complexity compared to generating all data at once.
2Manufacturing precision
If previous methods preserve many-to-many relationships in synthetic data, then data fidelity is improved, but privacy protection deteriorates
Solution Approach 1:
The patent applies parameter changes by using the bipartite joint degree distribution as an intermediate representation that captures relationship structure without preserving individual data points. By generating synthetic data from this aggregated statistical representation rather than from the original detailed data, the method preserves the essential many-to-many relationship patterns while removing identifiable information about specific individuals or entities, thus protecting privacy.
Solution Approach 2:
The patent introduces the bipartite joint degree distribution as an intermediary representation between the original data and the synthetic data generation process. This intermediary captures the essential relational structure (many-to-many relationships) in an aggregated form that does not expose individual data points, enabling synthetic data generation that preserves relationships while protecting privacy.
3Productivity
If previous methods generate synthetic data quickly, then productivity is improved, but scalability deteriorates
Solution Approach 1:
The patent applies universality by creating a general framework that can handle datasets of varying sizes and complexities through the same core process. The method uses the bipartite joint degree distribution which can be computed for any dataset size, and the generative model can be applied universally across different domains and data types, making it scalable rather than optimized for specific small-scale problems.
Data Source
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.