synSPORT

This pipeline generates and evaluates synthetic tabular sport sessions from a tabular data. The pipeline preserves participant-level fields such as athlete identity, age, gender, sport type, and training experience, then generates repeated synthetic session records for each participant. Results are saved as CSV metrics, synthetic sessions, plots, logs, and a local dashboard.

Dataset

The expected dataset is a tabular CSV located at:

data/tabular.csv

Default columns used by the pipeline:

• Athlete_ID: participant identifier.
• Training_Outcome: default target variable.
• Age, Gender, Sport_Type, Training_Experience_Years: default static participant fields.
• Remaining numeric and categorical columns are treated as session-level variables.

The default task is classification. You can change the target and task with:

TARGET=Training_Outcome TASK=classification bash run_test.sh

Models

The implemented models are:

• ctgan: GAN-based tabular generator for mixed numerical and categorical data.
• dpctgan: differentially private CTGAN for privacy-aware generation.
• pategan: privacy-preserving GAN using teacher-student aggregation.
• tabsyn: diffusion-based tabular generator using a latent denoising process.
• sdv_gaussian: Gaussian Copula statistical generator; fast baseline model.
• sdv_tvae: variational autoencoder for tabular data.
• sdv_par: sequential model for participant/session-style records.
• realtabformer: transformer-based tabular generator.
• bootstrap: lightweight resampling baseline for quick checks.

Install

Use the Python installation available in the working environment, then install the dependencies from the files in this folder:

cd synSPORT
python -m pip install --upgrade pip
python -m pip install -r requirements.txt

If your Python executable is not named python, set it explicitly:

PYTHON_BIN=/path/to/python bash run_test.sh

Quick Local Test

Use this first to confirm the pipeline works:

MODEL=sdv_gaussian \
SAMPLE_ROWS=120 \
SESSIONS_PER_PERSON=2 \
N_RUNS=1 \
EPOCHS=2 \
bash run_test.sh

Use MODELS=model_a,model_b only when a quick test should include more than one model.

Outputs are written to:

results/test/
logs/
dashboard/data/dashboard-data.js

Simulation Run

Use run_simulation.sh when you want to compare metrics over increasing model settings such as epoch steps. The script runs the simulation, merges the step/model outputs, and refreshes the dashboard data in one command.

Small local parallel example:

MODELS=sdv_gaussian,sdv_tvae \
SAMPLE_ROWS=120 \
SESSIONS_PER_PERSON=2 \
N_RUNS=2 \
STEPS=2 \
START_EPOCHS=2 \
EPOCH_STEP=2 \
MAX_WORKERS=2 \
WORKER_THREADS=1 \
OUTPUT_DIR=results/simulation_parallel_test \
LOG_FILE=logs/simulation_parallel_test.log \
bash run_simulation.sh

Larger production-style example:

MODELS=ctgan,dpctgan,pategan,tabsyn,sdv_gaussian,sdv_tvae,sdv_par,realtabformer \
SAMPLE_ROWS=0 \
SESSIONS_PER_PERSON=300 \
N_RUNS=5 \
STEPS=3 \
START_EPOCHS=50 \
EPOCH_STEP=50 \
MAX_WORKERS=2 \
WORKER_THREADS=1 \
bash run_simulation.sh

Increase MAX_WORKERS only if the machine has enough CPU and memory. Set MERGE_AFTER_RUN=0 only when you intentionally want to skip the automatic merge.

Simulation modes:

• SIMULATION_MODE=full: runs all configured models and steps. This is the default.
• SIMULATION_MODE=step: runs one model for one simulation step.
• SIMULATION_MODE=merge: merges completed step/model outputs and refreshes the dashboard data.

Dashboard

Start the dashboard with:

bash run_dashboard.sh

Open:

dashboard/production/index.html

To view a specific simulation output folder:

SIMULATION_DIR=results/simulation_parallel_test bash run_dashboard.sh

Batch Workflow

For larger runs, each model and epoch step can be executed as a separate job. Each job writes to a model-specific folder under results/simulation/.

Example single model-step job:

SIMULATION_MODE=step \
MODEL=ctgan \
STEP=1 \
EPOCHS=50 \
SAMPLE_ROWS=0 \
SESSIONS_PER_PERSON=300 \
N_RUNS=5 \
bash run_simulation.sh

The included pipeline.pbs file runs the same single model-step workflow with the variables supplied by the job environment.

After all separate model-step jobs finish, merge the outputs:

SIMULATION_MODE=merge bash run_simulation.sh

Then inspect the dashboard:

bash run_dashboard.sh

Files To Keep

The pipeline needs:

• data/
• models/
• dashboard/
• requirements.txt
• run_test.sh
• run_simulation.sh
• run_dashboard.sh
• pipeline.pbs
• README.md

The scripts create output folders such as results/ and logs/ when they run.

References

@article{challagundla2025synthetic,
  title={Synthetic Tabular Data Generation: A Comparative Survey for Modern Techniques},
  author={Challagundla, Raju and Dorodchi, Mohsen and Wang, Pu and Lee, Minwoo},
  journal={arXiv preprint arXiv:2507.11590},
  year={2025}
}