Database-Constrained GANs for Synthetic Data Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current approaches for generating synthetic data using generative adversarial networks (GANs) fail to consider database constraints, leading to inefficiencies and waste of computing resources due to the filtering out of invalid records.
Innovation Solution
Incorporate penalty components related to database constraints into the generator and discriminator loss functions of GANs, specifically using Tabular Generative Adversarial Networks (TGANs), to ensure generated synthetic data adheres to database constraints, thereby improving efficiency and reducing invalid records.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If traditional GAN approaches are used to generate synthetic data, then data availability is improved, but computing resource waste increases due to filtering invalid records
Solution Approach 1:
The patent applies preliminary action by incorporating database constraint checks into the generator loss function during the training phase. This allows the generator to learn and adhere to constraints before data generation, preventing the creation of invalid records in the first place rather than filtering them afterward, thus avoiding computing resource waste while maintaining data availability
2Quantity of substance
If traditional GAN approaches are used to generate synthetic data, then data availability is improved, but generation efficiency decreases due to post-generation filtering
Solution Approach 1:
The patent implements preliminary action by integrating constraint satisfaction mechanisms into the generator's learning process during training. The generator is penalized for producing records that violate database constraints, enabling it to generate valid synthetic data directly without requiring subsequent filtering steps, thereby improving generation efficiency while maintaining data availability
3Measurement precision
If database constraints are incorporated into loss functions, then data accuracy improves, but model complexity increases
Solution Approach 1:
The patent applies parameter changes by modifying the loss function parameters to include constraint satisfaction terms. Instead of changing the fundamental architecture of the GAN model, the approach adjusts the mathematical parameters (loss components) to guide the generator toward producing constraint-compliant synthetic data, thereby improving data accuracy while avoiding significant increases in model complexity
Data Source
AI summary
Disclosed herein are system, method, and computer program product embodiments for generating synthetic data records with database constraints using generative adversarial networks (GAN). The method can include training, by using a generator loss function, a generator neural network of a generator model of the GAN to generate a scaling factor and a cluster vector for a datum of a continuous variable of a continuous column of a data table, and a datum for a categorical variable of a categorical column of the data table. The generator loss function includes a penalty component determined based on a set of data constraints related to the continuous column or the categorical column.