p-mICE

Projected-multivariate Individual Conditional Expectation (p-mICE) is an XAI approach that extends ICE.
p-mICE aims to visualize changes in predicted values with the perturbation of explanatory variables while accounting for multivariate dependence.

Requirements

CentOS Linux release 8.1 (Confirmed)

Create conda environment

conda config --add channels conda-forge
conda create --name p_mice python=3.7.3
conda activate p_mice
conda install pandas=0.25.3
conda install numpy=1.16.0
conda install xgboost=0.82
conda install scikit-learn=0.21.2
conda install matplotlib=3.1.2
conda install seaborn=0.11.2
conda install jupyter=1.0.0

Tutorials

In brief, Tutorial.ipynb can be run on the Jupiter Notebook.

Import

import p_mice

p-mICE settings

pdc_params = p_mice.PdcParams(*model, *ar_feature_names, *intervention_variables, threshold=0.5, roughness=0.05, cont_index=None)

• *model: p-mICE can be applied to any binary classification ML model with predict_proba method.
• *ar_feature_names: 1d-array [str]. Feature names used in ML model. The array length should be the same as the number of explanatory variables in the model.
• *intervention_variables: List [str]. Two intervention variables corresponding to the x- and y-axes of the plot.
• threshold: [float]. Threshold value of the prediction probability for binary classification in the prediction model (default=0.5).
• roughness: [float] value in the range of [0, 1]. The smaller the roughness, the more finely the search is performed (default=0.05).
• cont_index: 1d-array [bool] to indicate continuous features. If None, all features are regarded as continuous variables (default=None).

p-mICE calculation for the individual record

pdc_params = p_mice.calc_p_mice(*pdc_params, *record, *ar_ref_X, ar_ref_y)

• *pdc_params: p-mICE setting.
• *record: 2d-array. Single record to explain.
• *ar_ref_X: 2d-array. Reference data to search neighbor records in p-mICE.
• same_label: [bool]. If True, reference data is restricted to records that have the same label as original predicted value (default=True).
• ar_ref_y: 1d-array [int]. Labels for ar_ref_X. If same_label is True, this option is required.
• ar_p_min: 1d-array [float]. Lower limits of variation for intervention variables. If None, lower limits in the ar_ref_X would be set (default=None).
• ar_p_max: 1d-array [float]. Upper limits of variation for intervention variables. If None, upper limits in the ar_ref_X would be set (default=None).
• n_neighbors: [int]. The number of neighbor records used in p-mICE (default=4).
• mean_seight: [str], choices=['simple', 'euclidean1', 'euclidean2', 'exp_euclidean1', 'exp_euclidean2']. The weights of neighbor records in p-mICE (default='exp_euclidean1').

Plot of p-mICE results

p_mice.plot_phase_diagrams(pdc_params)

• *pdc_params: Calculated p-mICE object.

License

This edition of p-mICE is for evaluation, learning, and non-profit academic research purposes only, and a license is needed for any other uses. Please send requests on license or questions to nakamura.kazuki.88m[at]st.kyoto-u.ac.jp