Tabular Data Reproduction With GFlowNets for Traceable Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing tabular data generation techniques, such as CTGAN and TabDDPM, generate good-quality data but lack the ability to analyze or interpret the generation process, leading to inefficiencies in high-value industries like healthcare and finance.
Innovation Solution
A method utilizing generative flow networks (GFlowNets) and a critic network with a processor that learns through GFlowNets and a critic network, preprocessing continuous variables into categorical variables and one-hot vectors, and using a Bernoulli distribution to sample trajectories, combined with Wasserstein GAN with gradient penalty (WGAN-GP) for reward-based learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If deep neural networks are used to generate tabular data (CTGAN, TabDDPM), then data quality is improved, but the ability to analyze and interpret the generation process deteriorates
Solution Approach 1:
The patent introduces GFlowNets as an intermediary framework between the generation process and the final output. This framework provides a structured flow network that mediates the transformation from input to output, enabling traceable generation processes while maintaining high data quality through the critic network's guidance.
Solution Approach 2:
The generation process is segmented into discrete flow networks with multiple layers and nodes. Each layer processes specific aspects of the data generation, allowing the complex generation process to be broken down into interpretable segments that can be analyzed individually while collectively producing high-quality tabular data.
2Adaptability or versatility
If continuous variables are used in training data, then data representation flexibility is improved, but processing complexity increases
Solution Approach 1:
The patent transforms continuous variables into discrete categorical variables through parameter changes. This conversion simplifies the processing complexity by converting continuous data into discrete states that can be more easily handled by the flow network, while still maintaining adaptability through the one-hot encoding scheme that preserves information about the original continuous values.
Data Source
AI summary
The present invention relates to an apparatus for reproducing tabular data. The apparatus for reproducing tabular data includes a generative flow networks (GFlowNets) learning network, a critic network, and a processor that performs learning through the GFlowNets learning network and the critic network, in which the processor performs learning of a policy network of GFlowNets and learning using real data, and performs learning of the critic network based on the real data and data generated by the GFlowNets.