TabularQGAN: a quantum generative model for tabular data synthesis

Area of Science:

• Quantum Computing
• Artificial Intelligence
• Data Science

Background:

• Synthetic data generation is crucial for augmenting scarce or private real-world datasets.
• Enterprise data is often tabular and heterogeneous, requiring specialized synthesis models.
• Existing quantum generative models are limited to homogeneous data types.

Purpose of the Study:

• To introduce a novel quantum generative model for synthesizing heterogeneous tabular data.
• To address the gap in quantum models capable of handling mixed data types.
• To evaluate the model's performance against state-of-the-art classical methods.

Main Methods:

• Proposed a quantum generative adversarial network (GAN) architecture with flexible data encoding.
• Developed a novel quantum circuit ansatz tailored for tabular data.
• Tested on MIMIC-III (healthcare) and Adult Census datasets using a noiseless statevector simulator.

Main Results:

• The quantum model achieved competitive and sometimes leading performance on the MIMIC-III dataset, based on SDMetrics.
• Demonstrated strong generalization capabilities with custom-designed metrics.
• Outperformed or matched leading classical models like CTGAN, CopulaGAN, VAE-GMM, and an LLM-based approach.

Conclusions:

• The proposed quantum generative model effectively synthesizes heterogeneous tabular data.
• This approach offers a promising alternative for privacy-preserving data augmentation in various industries.
• Further research can explore its application on larger datasets and with noise considerations.