Hybrid GAN - Copula Data Synthesizer

Hybrid Generative Adversarial Neural Network and Copula Data Synthesizer

Model Demo: [ https://drive.google.com/file/d/1fd-9cMRGWv_yN45iL_XvwTw8XLJSGH6C/view?usp=drive_link ]

A powerful Streamlit-based application that enables you to upload a real-world tabular dataset, preprocess it automatically, and generate high-quality synthetic datasets using Gaussian Copula and Generative Adversarial Network (GAN)–based synthesis models.

---

Overview

This project provides a hybrid framework for tabular data generation using both statistical and deep-learning-based generative models.

It combines the strengths of:

• Gaussian Copula models for learning statistical dependencies,
• Generative Adversarial Networks (CTGAN) for capturing complex, non-linear relationships.

The interface is built with Streamlit, providing an intuitive and interactive web-based environment for:

• Uploading datasets,
• Configuring synthesis parameters,
• Training models,
• Evaluating generated data quality,
• Downloading synthetic datasets.

---

Key Features

1. Dual Synthesis Engines

Choose between:

• Gaussian Copula: Fast, statistically driven model for smaller datasets.
• Generative Adversarial Network (GAN/CTGAN): Deep-learning approach for complex, large, or high-dimensional data.

2. Automated Preprocessing

• Detects column types automatically.
• Cleans missing values, encodes categories, parses datetimes.
• Handles high-cardinality columns with configurable Top-K mapping.
• Auto-converts incompatible types to SDV-supported formats.

3. Evaluation Metrics

After generating synthetic data, the app computes:

• Kolmogorov–Smirnov (KS) Statistic: For numeric distribution similarity.
• Jensen–Shannon (JS) Divergence: For categorical distributions.
• Correlation Matrix MSE: To assess structure preservation.
• ML Utility (AUC): Tests predictive similarity between real and synthetic data.
• Privacy Metrics (Memorization Check): Ensures synthetic data doesn’t memorize real rows.

4. Synthetic Data Export

• Instantly download your generated synthetic dataset as a CSV file.
• Columns and types aligned with your original dataset.

---

Tech Stack

Component	Technology
Frontend	Streamlit
Backend / Model Engine	SDV (Synthetic Data Vault)
Models Used	GaussianCopula, CTGAN
Language	Python 3.8+
Libraries	pandas, numpy, sdv, scipy, scikit-learn
Interface	Streamlit UI with custom CSS

---