QII

The qii_tool package is an implementation of the QII method proposed in the paper "Algorithmic Transparency via Quantitative Input Influence: Theory and Experiments with Learning Systems". The original paper discusses the transparency-privacy tradeoff, whereas this particular package only exploits its transparency aspect to be used as an influence measures for interpretable machine learning.

Install qii_tool on your system using:

pip install qii_tool

or clone the repository and run:

python setup.py bdist_wheel
python -m pip install dist/qii_tool-0.1.2-py3-none-any.whl

Following examples can be found in the experiments:

• iris dataset

• digits dataset

Extensions:

The package is implemented in such a way that it is easy to extend to user's need. Following are several examples:

• set evaluted_features in QII.compute() method to evaluate QII value for a selected set of features.
• QII.compute_unary_qii( si, S) can be used to compute Unary QII for feature s_i with respect to a feature set S.
• set pool to a specific distribution, e.g. all data points has feature sepal length > 4.6 cm, in QII.compute()

Credit:

• The code is adapted from Shayak Sen's version.
• For further request of case study or issues using the package, please contact Vinh (hovinh39@gmail.com).