Empirical copula-based data augmentation for mixed-type datasets: A robust approach for synthetic data generation

Abstract

Data augmentation is a critical technique for enhancing model performance in scenarios with limited, sparse, or imbalanced datasets. While existing methods often focus on homogeneous data types (e.g., continuous-only or categorical-only), real-world datasets frequently contain mixed data types (continuous, integer, and categorical), posing significant challenges for synthetic data generation. This article introduces a novel empirical copula-based framework for generating synthetic data that preserves both marginal and joint probability distributions and dependencies of mixed-type features. Our method addresses missing values, handles heterogeneous data through type-specific transformations, and introduces controlled noise to enhance diversity while maintaining statistical fidelity. We demonstrate the efficacy of this approach using synthetic and experimental benchmark datasets such as the Census Income and the Wisconsin Diagnostic Breast Cancer (WDBC) dataset, demonstrating its ability to generate realistic synthetic samples that retain the statistical properties of the original data. The proposed method is implemented in an open-source Python class, ensuring reproducibility and scalability.