def create_surrogate_model(explainer, type, max_vars, max_depth, **kwargs):
# utility function for model_surrogate(type=('tree', 'linear'))
y_hat = explainer.predict(
explainer.data) if explainer.y_hat is None else explainer.y_hat
# init a proper model
if explainer.model_type == 'regression':
if type == 'tree':
from sklearn.tree import DecisionTreeRegressor
surrogate_model = DecisionTreeRegressor(max_depth=max_depth,
**kwargs)
elif type == 'linear':
from sklearn.linear_model import LinearRegression
surrogate_model = LinearRegression(**kwargs)
elif explainer.model_type == 'classification':
y_hat = y_hat.round().astype(int) # modify prob to classes
if type == 'tree':
from sklearn.tree import DecisionTreeClassifier
surrogate_model = DecisionTreeClassifier(max_depth=max_depth,
**kwargs)
elif type == 'linear':
from sklearn.linear_model import LogisticRegression
surrogate_model = LogisticRegression(**kwargs)
else:
raise ValueError(
"'explainer.model_type' must be 'regression' or 'classification'")
# find the most important variables
if max_vars is not None and max_vars < len(explainer.data.columns):
# use model native feature importance if possible
if hasattr(explainer.model,
"feature_importances_"): # scikit, xgboost, lightgbm
# take last max_vars of the variables sorted from least to most important
vi = explainer.model.feature_importances_
variables = explainer.data.columns[np.argsort(vi)][-max_vars:]
else:
# calculate model_parts and take max_vars most important variables
vi = explainer.model_parts(N=500, B=15)
variables = vi.result.variable[~vi.result.variable.isin(
['_baseline_', '_full_model_'])].tail(max_vars)
X = explainer.data.loc[:, variables]
else:
variables = explainer.data.columns
X = explainer.data
surrogate_model.fit(X, y_hat)
# add additional attributes to the surrogate model object
surrogate_model.feature_names = variables.tolist()
surrogate_model.class_names = \
surrogate_model.classes_.astype(str).tolist() if hasattr(surrogate_model, 'classes_') else None
from .object import Explainer
surrogate_explainer = Explainer(surrogate_model,
X,
y_hat,
model_type=explainer.model_type,
verbose=False)
surrogate_model_performance = surrogate_explainer.model_performance()
surrogate_model.performance = surrogate_model_performance.result
# add the plot method to the surrogate model object
if type == 'tree':
# it uses feature_names and class_names if needed
import types
surrogate_model.plot = types.MethodType(plot_tree_custom,
surrogate_model)
# TODO: add plot method for the linear model
return surrogate_model
