Kolmogorov–Arnold Networks for Tabular Data Synthesis

KAN_synth is an open-source Python package for generating high-fidelity synthetic tabular data using Kolmogorov–Arnold Networks (KANs). It extends the original CTGAN and TVAE models from the Synthetic Data Vault (SDV) by replacing their MLP-based architectures with KAN-based components.

This project was developed for research (undergraduate thesis) and practical evaluation of KAN-based generative models, particularly in comparison to traditional GANs and VAEs on tabular data synthesis tasks.

Overview

The KAN_synth repository is structured into several components:

Models (models/)

KAN-based generative models for tabular data:

• KAN_CTGAN: Fully KAN-based implementation of CTGAN.
• HYBRID_KAN_CTGAN: CTGAN with only the first layer replaced by KAN.
• Disc_KAN_CTGAN: CTGAN with a KAN-based discriminator only.
• Gen_KAN_CTGAN: CTGAN with a KAN-based generator only.
• KAN_TVAE: Full KAN-based variant of TVAE.
• HYBRID_KAN_TVAE: TVAE where one intermediate layer is replaced by a KAN block.

Evaluation Scripts (benchmarks/)

Non-reusable scripts developed to:

• Train and evaluate each model on various datasets.
• Measure similarity and ML utility.
• Aggregate and visualize results.

Data Generation Scripts (data_gen/)

Standalone scripts for generating synthetic datasets using each model for internal benchmarking and experimentation purposes.

Utilities (utilities/)

Functions to compute:

• Overall similarity between real and synthetic datasets
• Model evaluation metrics (MAE, RMSE, R² for regression; Accuracy, F1, etc. for classification)
• Visualizations of ML utility scores across models

Tests (test/)

Unit tests for model importability, synthetic generation routines, and ML pipeline sanity checks using pytest.

The full model descriptions can be found in the thesis work at: [Not yet published].

Installation

Once the package is published on PyPI, it can be installed via pip:

pip install KAN-synth

All required packages are listed in requirements.txt, and will be automatically installed when using:

pip install -r requirements.txt

Since these models are based on the original CTGAN and TVAE architectures, the same data preprocessing principles apply:

• Continuous columns must be represented as floats.
• Discrete columns must be represented as integers or strings.
• The input dataset should not contain any missing values.

For additional details, refer to the original CTGAN repository: https://github.com/sdv-dev/CTGAN

Local Development Installation

if you'd like to install the package locally for development or experimentation:

git clone https://github.com/cris1618/KAN_synth.git
cd KAN_synth
pip install -e .

The -e flag installs the package in "editable" mode, meaning any changes you make to the code will immediately affect the installed version without needing to reinstall.

Usage Example

Here's a minimal working example to train a KAN-based synthesizer and generate synthetic tabular data using the KAN_CTGAN model.

from KAN_synth import KAN_CTGAN
import pandas as pd

# Load real tabular data (pandas DataFrame)
df = pd.read_csv("your_dataset.csv")

# Define the discrete columns (if any)
discrete_columns = ["column_a", "column_b"]  # Modify as needed

# Initialize the synthesizer
synthesizer = KAN_CTGAN(
    epochs=100, 
    verbose=True, 
    grid_size_gen=5,  
    spline_order_gen=3,
    grid_size_desc=5,
    spline_order_desc=3
)

# Train the model on your data
synthesizer.fit(df, discrete_columns)

# Sample 1000 synthetic rows
synthetic_data = synthesizer.sample(1000)

# Save or explore the results
synthetic_data.to_csv("synthetic_output.csv", index=False)

You can replace KAN_CTGAN with any of the other available models, such as:

• HYBRID_KAN_CTGAN
• Disc_KAN_CTGAN
• Gen_KAN_CTGAN
• KAN_TVAE
• HYBRID_KAN_TVAE

Each supports similar training and sampling APIs compatible with the original CTGAN and TVAE interfaces.