A tabular data generation framework guided by downstream tasks optimization.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Generative models are emerging for dataset extension.
• Existing models struggle with numerical constraints in tabular data generation.
• High-quality, constraint-satisfying tabular data is crucial for downstream applications.

Purpose of the Study:

• To propose a novel framework for tabular data generation guided by downstream task optimization (TDGGD).
• To address the limitations of current generative models in creating realistic tabular data with numerical constraints.
• To enhance the utility and prediction accuracy of generated tabular datasets.

Main Methods:

• Developed a tabular data generation framework (TDGGD) using diffusion models.
• Incorporated three indicators and gradient optimization within diffusion generation steps.
• Integrated downstream task information directly into the generative process, unlike traditional separate model strategies.

Main Results:

• TDGGD successfully generates tabular data that adheres to numerical column constraints.
• Experiments show enhanced prediction accuracy on real-world datasets with TDGGD.
• The framework demonstrates improved data utility over statistical fidelity alone.

Conclusions:

• TDGGD is the first diffusion model framework to integrate downstream task information for tabular data generation.
• The proposed method effectively increases data volume while improving predictive performance.
• TDGGD offers a promising approach for generating high-quality, application-specific synthetic tabular data.