DecomposableTabSyn

The implementation of the Waterloo CS848 project: Raw Table Synthesis through Decomposition

Overview

This project implements a framework for decomposing and synthesizing tabular data. The structure and functionality are designed to allow for extendability, enabling the integration of custom synthesizers and decomposers by inheriting basic classifiers provided in the synthesizers and decomposers modules.

Repository Structure

.
├── LICENSE
├── README.md
├── requirements.txt
├── setup.py
└── src
    ├── __init__.py
    ├── analysis_extract_metrics.py
    ├── analysis_separate.py
    ├── dataloader.py
    ├── decompose.py
    ├── decomposition.py
    ├── eval_synthesizer.py
    ├── eval_synthesizers.sh
    ├── evaluation.py
    ├── synthesis2.py
    ├── synthesizers
    └── utils.py

You can test the decomposition module independently:

python -m src.decomposition

Installation

Install the package in editable mode to enable development and testing:

pip install -e .

Running Experiments

The framework includes scripts to conduct five designed experiments and analyze their results.

Step 1: Execute the Experiments

Navigate to the src directory and run the evaluation script:

cd src
sh eval_synthesizers.sh

This script runs five predefined experiments using the implemented synthesizers.

Step 2: Extract Metrics and Integrate Results

Once the experiments are complete, extract the metrics:

python analysis_extract_metrics.py

This generates the file experiment_metrics_with_labels.csv that consolidates all experiment results.

Step 3: Visualize and Plot Results

Generate plots from the extracted metrics and save them into the images directory:

python analysis_separate.py

The resulting visualizations are stored as PDF files in the images folder.

Key Features

• Extendable synthesizers and decomposers: Create custom extensions by inheriting from base classifiers in the synthesizers and decomposers modules.
• Automated experiment execution: Run multiple predefined experiments using a single shell script.
• Integrated analysis: Extract metrics and visualize results for better interpretability.
• Modular design: Analyze, synthesize, and decompose tabular data through standalone modules.

Acknowledges

This project leverages functionality provided by the SDV (Synthetic Data Vault) library and SDMetrics library. The SDV library and SDMetrics offers comprehensive tools for synthetic data generation, including metrics for evaluating synthetic data quality and fidelity.