How orcads/data-001-impact-of-synthetic-data-under-data-scarcity works
This repository implements a healthcare tabular synthetic-data scarcity benchmark. Its completed Gaussian Copula workflow compares synthetic augmentation with real-only training across scarcity levels, repeated splits, and classifiers, evaluating on held-out real data. Diagnostics cover fidelity, validity, reliability, and proximity; a separate CTGAN workflow is operationally paused, with verification components retained. The predictive orchestration and generator wiring are supported by sampled provenance sources; broader capability links follow the README and tree, whose unsampled implementations limit call-level detail.
Main entry points: augmentation.py Gaussian comparison · baseline.py Real-only baseline · gaussian_diagnostics.py Gaussian diagnostics · ctgan_full_grid.py CTGAN full grid
Components 23
Data and scarcity
- Dataset loading
src/data/loader.py - Scarcity sampling
results/provenance/gaussian_predictive_execution_v1_sources/src/data/scarcity.py - Domain validation
src/data/validation.py - Feature preprocessing
src/data/preprocessing.py
Synthetic generation
- Gaussian Copula
results/provenance/gaussian_predictive_execution_v1_sources/src/generators/gaussian_copula.py - CTGAN generation
src/generators/ctgan.py - TVAE generator
src/generators/tvae.py
Study workflows
- Gaussian comparison
results/provenance/gaussian_predictive_execution_v1_sources/src/experiments/augmentation.py - Real-only baseline
src/experiments/baseline.py - Gaussian diagnostics
src/experiments/gaussian_diagnostics.py - CTGAN pilot
src/experiments/ctgan_pilot.py - CTGAN full grid
src/experiments/ctgan_full_grid.py - Logistic regression
src/models/logistic_regression.py - Random forest
src/models/random_forest.py
Evaluation and reporting
- Grid verification
src/evaluation/ctgan_full_grid_verifier.py - Predictive metrics
src/evaluation/predictive.py - Fidelity metrics
src/evaluation/fidelity.py - Proximity metrics
src/evaluation/privacy.py - Statistical summaries
src/evaluation/statistics.py - Result plots
src/visualization/plots.py - Exploratory analysis
src/visualization/eda.py
Other components
- Researcher
- Healthcare datasets
Connections 21
- Researcher → Gaussian comparison: runs comparison
- Healthcare datasets → Dataset loading: provides records
- Gaussian comparison → Scarcity sampling: samples training
- Gaussian comparison → Gaussian Copula: generates augmentation
- Gaussian comparison → Logistic regression: fits classifier
- Gaussian comparison → Random forest: fits classifier
- Gaussian comparison → Predictive metrics: scores predictions
- Real-only baseline → Scarcity sampling: samples training
- Real-only baseline → Logistic regression: fits classifier
- Real-only baseline → Random forest: fits classifier
- Real-only baseline → Predictive metrics: evaluates predictions
- Gaussian diagnostics → Gaussian Copula: generates samples
- Gaussian diagnostics → Domain validation: checks domain
- Gaussian diagnostics → Fidelity metrics: measures fidelity
- Gaussian diagnostics → Proximity metrics: measures proximity
- CTGAN pilot → CTGAN generation: runs pilot
- CTGAN full grid → CTGAN generation: dispatches conditions
- CTGAN full grid → Grid verification: verifies conditions
- Gaussian comparison → Statistical summaries: summarizes metrics
- Statistical summaries → Result plots: supplies summaries
- Dataset loading → Exploratory analysis: supports exploration