QTabGAN: A Hybrid Quantum-Classical GAN for Tabular Data Synthesis

Synthesizing realistic tabular data is challenging due to heterogeneous feature types and high dimensionality. We introduce QTabGAN, a hybrid quantum-classical generative adversarial framework for tabular data synthesis. QTabGAN is especially designed for settings where real data are scarce or restricted by privacy constraints. The model exploits the expressive power of quantum circuits to learn complex data distributions, which are then mapped to tabular features using classical neural networks. We evaluate QTabGAN on multiple classification and regression datasets and benchmark it against leading state-of-the-art generative models. Experiments show that QTabGAN achieves up to 54.07% improvement across various classification datasets and evaluation metrics, thus establishing a scalable quantum approach to tabular data synthesis and highlighting its potential for quantum-assisted generative modelling.

💡 Research Summary

QTabGAN introduces a hybrid quantum‑classical generative adversarial network tailored for realistic tabular data synthesis. The generator’s core is a variational quantum circuit (VQC) composed of n qubits and L layers; each layer applies a Hadamard initialization, parameterized RY and RZ rotations, and a circular CNOT entanglement pattern. This design exploits quantum superposition and entanglement to produce a 2ⁿ‑dimensional probability distribution that captures complex inter‑feature correlations often missed by classical generators.

After the VQC is trained, a quantum sampler measures the circuit in the computational basis many times (N > 2ⁿ) to obtain an empirical probability vector pθ. Because pθ does not directly correspond to tabular features, a classical feed‑forward neural network—referred to as the Classical Mapper (CLMapper)—transforms the probability vector into synthetic samples. The mapper handles both continuous and categorical variables, optionally conditioning on class labels to enable class‑conditional generation.

The discriminator is a conventional multilayer perceptron that distinguishes real from synthetic rows. Training follows the standard GAN minimax objective, but the quantum parameters are updated via the parameter‑shift rule combined with a classical optimizer such as Adam. This creates a fully integrated hybrid learning loop where quantum and classical components co‑evolve.

The authors evaluate QTabGAN on six publicly available tabular datasets spanning classification and regression tasks (e.g., Adult, Credit Card, Higgs, Bike Sharing). They assess statistical fidelity using KL‑divergence and Jensen‑Shannon distance, and downstream utility by training downstream models (logistic regression, XGBoost, neural nets) on synthetic data and measuring F1‑score, AUROC, or RMSE. Across all benchmarks, QTabGAN outperforms state‑of‑the‑art classical tabular GANs such as CTGAN, TableGAN, CTAB‑GAN+, and CasT‑GAN. Reported improvements reach up to 54.07% on certain metrics, with an average gain of roughly 30% over baselines. Notably, when the amount of real data is reduced to as little as 1 % of the original size, the quantum generator still maintains superior performance, suggesting that the high‑dimensional quantum latent space mitigates over‑fitting and enhances generalisation.

A systematic ablation study examines the impact of circuit depth (L) and qubit count (n). The best trade‑off occurs with L ≈ 4–6 layers and n = 8–10 qubits, which provide sufficient expressive power while keeping gate count manageable for noisy intermediate‑scale quantum (NISQ) devices. Increasing n beyond 12 dramatically raises simulation cost and would likely exacerbate hardware noise, limiting practical scalability at present.

The paper acknowledges several limitations. All experiments are conducted on quantum simulators; real‑hardware validation under realistic noise models remains future work. Moreover, privacy‑preserving mechanisms such as differential privacy are not integrated, even though the authors motivate the approach for privacy‑sensitive scenarios.

Future directions include implementing error‑mitigation and noise‑robust circuit designs, coupling the framework with formal privacy guarantees, and testing on emerging NISQ processors to assess real‑world feasibility.

In summary, QTabGAN demonstrates that a variational quantum circuit can serve as a powerful latent‑space generator for tabular data, and that coupling it with a classical mapper yields synthetic datasets of higher fidelity than existing classical methods. The work opens a promising avenue for quantum‑assisted data generation, especially in regimes where data are scarce or privacy constraints limit direct data sharing.