Abstract Deep learning for tabular data presents unique challenges compared to structured data like images or text due to the diversity in feature types, complex relationships between variables, and the frequent occurrence of missing values. Traditional machine learning models like decision trees, random forests, and gradient boosting methods have been more successful on tabular data compared to neural networks, which often require more effort due to weak inductive bias, difficulty capturing heterogeneous feature interactions, and sensitivity to limited or imbalanced samples; but deep learning techniques are evolving to address these limitations. In the literature, TabNet is a deep learning architecture designed specifically for tabular data. It employs attention mechanisms to focus on the most relevant features at each decision step, mimicking how tree-based models like gradient boosting trees work. This research introduces an enhanced TabNet-based model, called CT-TabGAN (Conditional-TabNet GAN), built on the foundation of Tabular Generative Adversarial Network (TGAN) architectures to effectively address class imbalance in tabular datasets. To achieve this, the generator is redesigned to incorporate TabNet-aware feature structuring, allowing for seamless integration into the TabNet training pipeline. Furthermore, the model incorporates an improved conditional sampling module that more effectively preserves the complex dependencies between high-dimensional categorical and continuous features. On benchmark datasets, the proposed CT-TabGAN model consistently outperforms traditional oversampling techniques, achieving F1-scores of 0.93 for Credit Card Fraud Detection, 0.82 for Diabetes Prediction, and 0.92 for Customer Churn while also attaining higher AUC-ROC values of 0.96, 0.88, and 0.95, respectively, demonstrating its superior ability to handle class imbalance in real-world tabular datasets.
Darwish et al. (Tue,) studied this question.