synthetic-data

Inspired by sklearn.datasets.make_classification, which in turn is based on work for the NIPS 2003 feature selection challenge [1] - targeting linear classifiers. Here the focus is on generating more complex, nonlinear datasets appropriate for use with deep learning/black box models which 'need' nonlinearity - otherwise you would/should use a simpler model.

Approach

Ideally, the method would provide a concise specification to generate tabular data with sensible defaults. The specification should provide knobs that the end user can dial up or down to see it's downstream impact.

Copulas are a model for specifying the joint probability p(x1, x2, ..., xn) given a correlation structure along with specifications for the marginal distribution of each feature. The current implementation uses a multivariate normal distribution with specified covariance matrix. Future work can expand this choice to other multivariate distributions.

Parameters

Getting Started

Local Installation

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
pip install -e .

Tests

Test/Lint Dependencies

pip install -r requirements-test.txt

To run tests:

make test_local

Pre-Commit

To install pre-commit hooks, run the following commands:

pre-commit install
pre-commit run

Referencing this library

If you use this library in your work, please cite our paper:

@inproceedings{barr:2020,
  author    = {Brian Barr and Ke Xu and Claudio Silva and Enrico Bertini and Robert Reilly and  C. Bayan Bruss and Jason D. Wittenbach},
  title     = {{Towards Ground Truth Explainability on Tabular Data}},
  year      = {2020},
  maintitle = {International Conference on Machine Learning},
  booktitle = {2020 ICML Workshop on Human Interpretability in Machine Learning (WHI 2020)},
  date = {2020-07-17},
  pages = {362-367},
}

Notes

If you have tabular data, and want to fit a copula from it, consider this python library: copulas Quick visual tutorial of copulas and probability integral transform.

To run the examples, you should run:

$ python -m pip install pandas pytest pytest-cov seaborn shap tensorflow "DataProfiler[full]"

References

[1] Guyon, “Design of experiments for the NIPS 2003 variable selection benchmark”, 2003.