Interpretation

Several motivations exist for having an interpretability layer for ML models, including: a) being able to justify predictions, b) being able to diagnose vulnerabilities of models, to be able to c) improve them, and c) to understand better our reality, as when algorithms discover relevant patterns or relations, we can build upon them to generate new knowledge. (source)

The interpretation module provides various tools for understanding your models. These tools are Model-Agnostic, meaning they do not depend on the specific model used. Here is a list of methods available in this module:

Partial Dependence Plot (PDP),
Individual Condition Expectation (ICE),
Accumulated Local Effects (ALE),
Diverse Counterfactual Explanations (DiCE),
Local Interpretable Model-Agnostic Explanations (LIME).

The interpretation module consists of three classes: ExplanationPlot, DiCEExplanation, LimeExplanation.

PDP, ICE and ALE

ExplanationPlot is a class for creating a plot for interpretation. Partial Dependence Plot (PDP), Individual Condition Expectation (ICE), Accumulated Local Effects (ALE) are supported. You can select the method by changing the method parameter: pdp, ice or ale.

You need x, a fitted model and a list of features (features to interpret) to create a plot for interpretation.

import pandas as pd
from sklearn.ensemble import GradientBoostingRegressor
from insolver.interpretation import ExplanationPlot, DiCEExplanation, LimeExplanation

dataset = pd.read_csv(...)
x, y = dataset.drop(['y'], axis=1), dataset['y']
clr = GradientBoostingRegressor().fit(x, y)
features = x.columns

When all the data and model are set up, you just need to create a class instance and call the plot() method.

ep = ExplanationPlot(method='pdp', x=x, estimator=clr, features=features)
ep.plot()

You can change plot by setting different parameters in the plot() method.

ep = ExplanationPlot(method = 'ice', x=x, estimator=clr, features=features)
ep.plot(ice_lines_kw={"color": "tab:blue", "alpha": 0.2, "linewidth": 0.2},
        pd_line_kw={"color": "tab:orange", "linestyle": "--"})

DiCE

DiCEExplanation is a class for creating Diverse Counterfactual Explanations (DiCE). It uses dice_ml package. It supports only sklearn models.

You need x and y OR dataset and outcome name, a fitted model to create an explanation.

import pandas as pd
from sklearn.ensemble import GradientBoostingRegressor

dataset = pd.read_csv(...)
x, y = dataset.drop(['y'], axis=1), dataset['y']
clr = GradientBoostingRegressor().fit(x, y)
features = x.columns

Create a class instance and call the show_explanation() method with the data instance you need to explain. model_type patameter can be regressor or classifier.

de = DiCEExplanation(estimator=clr, model_type='regressor', x=x, y=y, continuous_features=features)
de.show_explanation(instance=x[0:1], total_CFs=3, desired_range=[50, 55])

You can also change the method parameter. Values genetic, random and kdtree are supported. show_explanation() has different parameters to specify an explanation: features_to_vary, permitted_range, result_type, return_json

de = DiCEExplanation(estimator=clr, model_type='regressor', x=dataset, #set x as dataset and outcome_name
                     continuous_features=features, 
                     outcome_name=['y'])

de.show_explanation(instance=x[0:1], total_CFs=3, desired_range=[50, 55], 
                    features_to_vary=['X1', 'X2', 'X3'],
                    permitted_range={'X3': [378, 400]})

LIME

LimeExplanation is a class for creating Local Interpretable Model-Agnostic Explanations (LIME). Uses lime package.

You need x and a fitted model to create an interpretation.

import pandas as pd
from sklearn.ensemble import GradientBoostingRegressor

dataset = pd.read_csv(...)
x, y = dataset.drop(['y'], axis=1), dataset['y']
clr = GradientBoostingRegressor().fit(x, y)
features = x.columns

Create a class instance and call the show_explanation() method with the data instance you need to explain. mode patameter can be regression or classification. Categorical features are set in the categorical_features parameter.

le = LimeExplanation(estimator=clr, x=x, mode='regression', features_names=features)
le.show_explanation(instance=x.iloc[1])

You can change the result_type parameter, values html, map, list, file, show_in_notebook are supported.

le.show_explanation(instance=x.iloc[1], result_type='html')