Generating High-Diversity Synthetic Tabular Data via a Less-Constrained Prior

This is the official implementation of the paper
"Generating High-Diversity Synthetic Tabular Data via a Less-Constrained Prior" (IJACI-ECAI 2026).

---

📦 Environment Setup

We recommend using a virtual environment.

Install the required dependencies using:

pip install -r requirements.txt

---

🔑 Weights & Biases (wandb) Setup

This codebase uses Weights & Biases (wandb) for experiment tracking.

1. Create a wandb account at https://wandb.ai
2. Log in from the terminal:

wandb login

3. In main.py or inference.py, set your wandb information at the top of the file:

project = "YOUR_PROJECT_NAME"
entity  = "YOUR_WANDB_USERNAME"

---

🏋️‍♂️ Training / Inference

This repository provides implementations of our proposed method (SPT) as well as baseline methods such as CTGAN, TVAE, TabMT, TabDDPM, STaSy, TabSyn, and AutoDiff.

The overall workflow is as follows:

1. Prepare a wandb account and configure it in the code
2. Move to the directory of the desired model (e.g., SPT, TabSyn)
3. Train the model using main.py
4. Generate synthetic samples using inference.py

---

Step 1: Move to the Model Directory

First, navigate to the directory corresponding to the baseline or method you wish to execute. For example:

cd SPT
# or
cd baseline/TabSyn
# or
cd baseline/TabDDPM

---

Step 2: Training

To train a model, run the following command:

python main.py --dataset [NAME_OF_DATASET] [ADDITIONAL_ARGS]

Here, [ADDITIONAL_ARGS] corresponds to model-specific hyperparameters, which are summarized below.

Required arguments by model

Model	Required Arguments
SPT	dataset, var, d_token, denoising_dim, lr1, batch_size1, batch_size2, max_beta
TabSyn	dataset, d_token, denoising_dim, lr1, batch_size1, batch_size2, max_beta
TabDDPM	dataset, weight_decay, lr, dim_embed, batch_size, num_layers
CTGAN	dataset, latent_dim, batch_size, epochs, generator_dim, discriminator_dim
TVAE	dataset, latent_dim, batch_size, epochs, loss_factor
TabMT	dataset, dim_transformer, num_transformer_heads, num_transformer_layer, batch_size, epochs, max_clusters
STaSy	dataset, model.sigma_min, model.sigma_max, optim.lr, model.beta0, model.alpha0
AutoDiff	dataset, d_token, denoising_dim, lr1, batch_size1, batch_size2

Notes

1. Supported datasets for --dataset are:

abalone, adult, anuran, banknote, breast, concrete,
kings, letter, loan, redwine, shoppers, whitewine

2. To ensure reproducibility, you can fix the random seed during training by adding the --seed argument.

3. Below is an example command for training SPT on the redwine dataset with a fixed seed:

python main.py --dataset redwine --var 0.1 --d_token 4 --denoising_dim 1024 \
               --lr1 0.001 --batch_size1 512 --batch_size2 512 --max_beta 0.01 --seed 0

---

Step 3: Inference (Synthetic Data Generation)

After training, synthetic samples can be generated using the following command:

python inference.py --dataset [NAME_OF_DATASET] [ADDITIONAL_ARGS]

Here, [ADDITIONAL_ARGS] must be identical to those used during training, as they are required to correctly load the trained model checkpoint.

Notes

1. The same set of hyperparameters used in the training phase must be provided during inference.

2. During inference, the argument --ver is used to specify the experiment version stored in wandb.

   • --ver controls which trained checkpoint is loaded and does not need to match seed

3. Below is an example command for generating synthetic data using SPT trained on the redwine dataset:

python inference.py --dataset redwine --var 0.1 --d_token 4 --denoising_dim 1024 \
                     --lr1 0.001 --batch_size1 512 --batch_size2 512 --max_beta 0.01 --ver 0

This command loads the trained SPT model corresponding to version ver = 0 from wandb and generates synthetic samples for the redwine dataset.

---

⚙ Example & Reproducibility

Quick Start with Jupyter Notebook

We provide a comprehensive example to ensure easy reproducibility of our results.

• File Location: SPT/example.ipynb
• Content: This notebook contains the entire pipeline for the anuran dataset, including:

1. Training: Step-by-step model training using the SPT method.
2. Inference: Generating synthetic tabular data from the trained model.
3. Evaluation: Measuring the quality of synthetic data using all metrics (GoF, MMD, PCD, Coverage, etc.).

You can run this notebook to verify the performance and understand the workflow of our proposed method.

---

Cite