Changshuo Liu
Changshuo Liu has a total of 1 co-authors. Here are the top co-authors by number of shared publications:
- Canyao Liu
Yale School of Management, Yale University, New Haven, CT 06511, USA
1 shared publication
Latest Publications
- Open Access
- 14 November 2024
- Published by Association for Computing Machinery (ACM)
- p. 547-554
Entity-based Financial Tabular Data Synthesis with Diffusion Models
In the rapidly evolving financial industry, the adoption of synthetic tabular data is on the rise to augment scarce data and facilitate data sharing. Existing synthetic tabular data generation methods often overlook a crucial aspect of financial datasets: the presence of a special column representing entities, such as companies or individuals. Treating this column as a mere categorical feature has significant drawbacks, including limiting the model to the original dataset’s entities, increasing privacy risks, and lacking efficient methods for generating data conditioned on specific entities. We introduce EntTabDiff, a diffusion model-based method that addresses these challenges by synthesizing tabular data conditioned on entities. EntTabDiff comprises two components: one that learns the distribution of entities to generate synthetic ones, and another that synthesizes tabular data based on these entities. We demonstrate the effectiveness of EntTabDiff using three financial datasets. Experimental results show that EntTabDiff synthesizes data with high utility and improved fidelity.
- 14 November 2024
- Published by Association for Computing Machinery (ACM)
- p. 547-554
Open AccessEntity-based Financial Tabular Data Synthesis with Diffusion Models
In the rapidly evolving financial industry, the adoption of synthetic tabular data is on the rise to augment scarce data and facilitate data sharing. Existing synthetic tabular data generation methods often overlook a crucial aspect of financial datasets: the presence of a special column representing entities, such as companies or individuals. Treating this column as a mere categorical feature has significant drawbacks, including limiting the model to the original dataset’s entities, increasing privacy risks, and lacking efficient methods for generating data conditioned on specific entities. We introduce EntTabDiff, a diffusion model-based method that addresses these challenges by synthesizing tabular data conditioned on entities. EntTabDiff comprises two components: one that learns the distribution of entities to generate synthetic ones, and another that synthesizes tabular data based on these entities. We demonstrate the effectiveness of EntTabDiff using three financial datasets. Experimental results show that EntTabDiff synthesizes data with high utility and improved fidelity.