Evaluating Normalizing Flow Model Variants for Supervised Tabular Data: A Comparative Analysis Using the Data Accuracy Score

In machine learning and data analytics, data is often organized in a tabular form because it is structured and easy to interpret. However, tabular data poses challenges for training machine learning models, such as overfitting and generalization issues, due to its limited availability. Data augmentation is a technique that addresses these challenges and improves model performance, especially with limited datasets. This paper explores normalizing flow models as a data augmentation technique for generating synthetic supervised tabular data. Evaluating the quality of such data is challenging because it lacks the visual features unlike images. Therefore, this paper introduces a novel evaluation method that assesses the performance of a classifier or regression model trained on the generated synthetic data. This method provides an indirect measure of the data’s accuracy through its effectiveness in supporting downstream tasks. This measure, called the Data Quality Score, is used to compare the performance of various normalizing flow models on supervised tabular data. The findings suggest that normalizing flows are a viable approach for data augmentation in supervised learning tasks involving tabular data and can generate synthetic data with a Data Quality Score of up to 99%.