FakeTables: Using GANs to Generate Functional Dependency Preserving Tables with Bounded Real Data

Haipeng Chen, Sushil Jajodia, Jing Liu, Noseong Park, Vadim Sokolov, V. S. Subrahmanian

Introduction

In sectors like healthcare and government, privacy concerns often restrict the release of comprehensive datasets, leading to a state of "data poverty." For instance, the US Census Bureau might only release a 1% sample of its records. Training machine learning models on such limited data often results in poor generalization and inaccuracy. Standard data augmentation techniques, such as SMOTE or traditional Generative Adversarial Networks (GANs), frequently fail to capture global statistical properties or respect the logical constraints inherent in relational data, known as Functional Dependencies (FDs).

Methodology: The ITS-GAN Framework

To address these challenges, we introduce ITS-GAN (Incomplete Table Synthesis GAN). This framework is designed to augment small, incomplete tabular datasets by ensuring synthetic records adhere to both global statistical averages and strict relational rules.

The ITS-GAN architecture incorporates pre-trained autoencoders to enforce functional dependencies and uses global statistics to guide the generator.

The core innovations of ITS-GAN include:

  1. FD Modeling with Autoencoders: Since discrete logical constraints like $A \rightarrow B$ are difficult to integrate into neural networks directly, ITS-GAN uses dedicated autoencoders for each FD. These models learn to reconstruct dependent values; high reconstruction error indicates a violation of the relational rule.
  2. Constraint-Aware Loss Function: The generator's loss is augmented with an $\alpha$-term to match known global column-wise averages and a $\beta$-term to penalize FD violations detected by the autoencoders.
  3. Enhanced Discriminator: The discriminator receives FD error terms as additional inputs, forcing the generator to produce logically consistent data to avoid detection.

Experimental Setup

The framework was evaluated on two real-world datasets:

  • US Census Bureau Data: Used to predict "hours-per-week" with an enforced FD: (gender, relationship) $\rightarrow$ marital-status.
  • US Bureau of Transportation Statistics (BTS): Used to predict "ticket price" with an enforced FD: (source/destination airports, flown distance) $\rightarrow$ overall distance.

In both scenarios, "data poverty" was simulated by providing only 1% or 5% of the original records for training.

Results and Discussion

ITS-GAN consistently outperformed state-of-the-art methods like TableGAN and SMOTE in classification accuracy. A key finding was that a model trained on a 1% sample augmented by ITS-GAN often matched or exceeded the performance of a model trained on a real 15% sample, effectively "filling in" the missing distribution.

Cumulative Distribution Function (CDF) plots for various features across the BTS and Census datasets, comparing ITS-GAN against baselines and the full table ground truth.

As shown in the CDF plots, ITS-GAN's synthetic data follows the "true" full-table distribution much more closely than competitors, which tend to overfit the small released sub-table. By turning discrete logical constraints into differentiable error terms, ITS-GAN successfully balances visual similarity, global statistical alignment, and relational integrity.

Conclusion

ITS-GAN provides a robust solution for generating high-fidelity synthetic relational data under conditions of data scarcity. While it currently requires pre-defined FDs and uses basic categorical encoding, it represents a significant step forward in functional dependency-preserving data synthesis. Future work will explore handling unknown or "soft" dependencies and more complex categorical embeddings.

Summary updated

Abstract

In many cases, an organization wishes to release some data, but is restricted in the amount of data to be released due to legal, privacy and other concerns. For instance, the US Census Bureau releases only 1% of its table of records every year, along with statistics about the entire table. However, the machine learning (ML) models trained on the released sub-table are usually sub-optimal. In this paper, our goal is to find a way to augment the sub-table by generating a synthetic table from the released sub-table, under the constraints that the generated synthetic table (i) has similar statistics as the entire table, and (ii) preserves the functional dependencies of the released sub-table. We propose a novel generative adversarial network framework called ITS-GAN, where both the generator and the discriminator are specifically designed to satisfy these two constraints. By evaluating the augmentation performance of ITS-GAN on two representative datasets, the US Census Bureau data and US Bureau of Transportation Statistics (BTS) data, we show that ITS-GAN yields high quality classification results, and significantly outperforms various state-of-the-art data augmentation approaches.
BibTeX
@inproceedings{chen2019faketables,
  author = {Haipeng Chen and Sushil Jajodia and Jing Liu and Noseong Park and Vadim Sokolov and V. S. Subrahmanian},
  title = {{FakeTables: Using GANs to Generate Functional Dependency Preserving Tables with Bounded Real Data}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.24963/ijcai.2019/287},
}
Figure