A Novel GAN Approach to Augment Limited Tabular Data for Short-Term Substance Use Prediction

Nguyen Thach, Patrick Habecker, Bergen Johnston, Lillianna Cervantes, Anika Eisenbraun, Alex Mason, Kimberly Tyler, Bilal Khan, Hau Chan

Abstract

Predicting substance use behavior is a critical component of public health strategy, yet researchers often face a "data desert" characterized by High-Dimension Low-Sample-Size (HDLSS) datasets. Standard generative models frequently fail in this context, producing synthetic data that violates logical survey constraints or fails to improve downstream classification. This paper introduces a specialized Generative Adversarial Network (GAN) architecture designed to augment small, high-dimensional survey datasets. By integrating survey "skip logic" constraints and a global feature selection mechanism, the proposed model generates high-quality synthetic samples that significantly improve the accuracy of short-term substance use predictions, outperforming existing state-of-the-art generative models.

Introduction

Accurate prediction of substance use behavior is essential for the efficient allocation of healthcare resources and the implementation of timely interventions. However, collecting data from Persons Who Use Drugs (PWUDs) is fraught with challenges, including high costs and recruitment difficulties. This often results in datasets where the number of features (variables) significantly exceeds the number of participants—a condition known as High-Dimension Low-Sample-Size (HDLSS).

Furthermore, survey-based data is inherently structured by "skip logic" (e.g., a respondent who has never used a substance should not be asked about the frequency of its use). Conventional generative models, such as standard GANs, typically ignore these logical dependencies, resulting in synthetic data that is internally contradictory. This research addresses these dual challenges by developing a GAN capable of producing logically consistent and high-utility synthetic data for substance use forecasting.

Methodology

The researchers developed a modified version of the Conditional Tabular GAN (CTGAN), introducing several architectural innovations tailored for survey-based HDLSS data:

1. Enforcing Skip Logic

The model mathematically defines skip constraints and integrates them directly into the GAN’s conditional vector. This ensures the generator is restricted to producing samples that adhere to the internal logic of the original survey instrument. Rather than merely acting as a post-processing filter, this logic serves as a regularizer during training.

2. Auxiliary Classifier with Feature Selection

To mitigate the risk of overfitting inherent in small datasets, the authors added an auxiliary classifier ($C$). This component utilizes a Weight Predictor Network (WPFS) to perform global feature selection. By identifying and focusing on the most relevant variables for the target outcome during the training process, the GAN avoids being misled by noise in the high-dimensional feature space.

3. Target-Prioritized Sampling

In HDLSS scenarios, specific conditions or minority classes may be absent from individual training batches. The model addresses this by increasing the sample size of the conditional vector and prioritizing the sampling of the target variable (substance use). This ensures the model learns the critical correlations necessary for predictive tasks.

4. Downstream Loss Integration

The architecture incorporates a "downstream loss" term. The generator is penalized if the auxiliary classifier cannot accurately predict the label of the synthetic data it produces, forcing the generator to create data that is not only realistic but also useful for training predictive models.

Experimental Setup

The model was evaluated using a unique longitudinal survey dataset involving 258 PWUDs in the U.S. Great Plains. The raw data included 151 features, which expanded to over 600 after one-hot encoding. The research focused on two primary tasks related to the use of marijuana, methamphetamine, amphetamines, and cocaine over a 12-month period:

  • Task A (Binary): Predicting whether a participant's usage would increase.
  • Task B (Multiclass): Predicting the specific frequency of future usage.

Key Results and Discussion

The proposed GAN was benchmarked against several state-of-the-art (SOTA) models, including CTGAN, TVAE, TableGAN, and TabDDPM. The results demonstrated significant improvements across multiple metrics:

  • Predictive Performance: Augmenting real training data with synthetic samples from the proposed GAN improved the Area Under the Receiver Operating Characteristic (AUROC) by up to 13.4% for Task A and 15.8% for Task B.
  • Logical Consistency: The model achieved a "Conflict" score of 0.196, substantially lower than the next best model (TVAE at 0.297). This indicates a much higher adherence to survey skip logic.
  • Training Stability: Ablation studies confirmed that enforcing skip logic stabilized the training phase, leading to lower loss values and more efficient convergence compared to versions of the model without these constraints.

One of the most significant insights from this study is the distinction between "compatibility" (how much synthetic data resembles real data) and "utility" (how much synthetic data improves a classifier). While many SOTA models can produce data that looks realistic, the proposed model is uniquely effective at improving classifier performance in data-scarce environments.

Conclusion

This research demonstrates that deep generative models can be successfully adapted for HDLSS survey data by incorporating domain-specific constraints and feature selection mechanisms. While the study is limited by its small, geographically specific sample size and the inherent biases of self-reported data, it provides a robust framework for data augmentation in sensitive healthcare domains. Future work will explore the application of these techniques to other fields with strict logical constraints, such as financial auditing and medical diagnostics.

Summary updated

BibTeX
@inproceedings{thach2024novel,
  author = {Nguyen Thach and Patrick Habecker and Bergen Johnston and Lillianna Cervantes and Anika Eisenbraun and Alex Mason and Kimberly Tyler and Bilal Khan and Hau Chan},
  title = {{A Novel GAN Approach to Augment Limited Tabular Data for Short-Term Substance Use Prediction}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
  url = {https://openreview.net/forum?id=yuljttLRJe},
}
Figure