GEM-T: Generative Tabular Data via Fitting Moments.

Area of Science:

• Data Science
• Machine Learning
• Artificial Intelligence

Background:

• Tabular data is prevalent but challenging for generative models, especially with limited or sensitive information.
• Existing generative models struggle with the heterogeneous nature and complex interactions within tabular datasets.

Purpose of the Study:

• To introduce GEM-T (Generative Entropy Maximization for Tables), a novel approach for synthetic tabular data generation.
• To demonstrate GEM-T's effectiveness in capturing higher-order column interactions and handling diverse data types.

Main Methods:

• The study utilizes the principle of maximum entropy (MaxEnt) to develop the GEM-T model.
• GEM-T directly models n-th order interactions (e.g., pairwise, third-order) within the training data's columns.
• Data preprocessing involves appropriate transformations to reveal low-dimensional correlations.

Main Results:

• GEM-T matches or surpasses state-of-the-art deep neural network approaches on 68% (23 of 34) of diverse public datasets.
• The model requires significantly fewer trainable parameters compared to deep learning methods.
• GEM-T effectively handles heterogeneous data types (continuous, discrete, categorical) and lack of local structure.

Conclusions:

• GEM-T offers a promising, lightweight, and high-performance alternative for generating synthetic structured data.
• The maximum entropy principle provides a robust framework for modeling complex interactions in tabular data.
• The findings suggest that valuable information in tabular data often resides in low-dimensional correlations, accessible through appropriate transformations.