def test_verify_linear_model_coefficient_explanation(self):
# Validate our explainer against an explainable linear model
X, y = shap.datasets.adult()
x_train, x_test, y_train, y_test = train_test_split(X,
y,
test_size=0.2,
random_state=7)
# Fit a logistic regression classifier
model = create_sklearn_logistic_regressor(x_train, y_train)
# Create tabular explainer
exp = TabularExplainer(model,
x_train,
features=list(range(x_train.shape[1])))
test_logger.info(
"Running explain model for test_verify_linear_model_coefficient_explanation"
)
# Validate evaluation sampling
policy = {
ExplainParams.SAMPLING_POLICY:
SamplingPolicy(allow_eval_sampling=True)
}
explanation = exp.explain_global(x_test, **policy)
mean_train = np.mean(x_train.values, axis=0)
# Retrieve the model coefficients
coefficients = model.coef_[0]
# Normalize the coefficients by mean for a rough ground-truth of importance
norm_coeff = mean_train * coefficients
# order coefficients by importance
norm_coeff_imp = np.abs(norm_coeff).argsort()[..., ::-1]
# Calculate the correlation
validate_correlation(explanation.global_importance_rank,
norm_coeff_imp, 0.76)
def tabular_explainer_imp(model, x_train, x_test, allow_eval_sampling=True):
# Create tabular explainer
exp = TabularExplainer(model,
x_train,
features=list(range(x_train.shape[1])))
# Validate evaluation sampling
policy = {
ExplainParams.SAMPLING_POLICY:
SamplingPolicy(allow_eval_sampling=allow_eval_sampling)
}
explanation = exp.explain_global(x_test, **policy)
return explanation.global_importance_rank
def test_explain_model_sparse_tree(self, tabular_explainer):
X, y = retrieve_dataset('a1a.svmlight')
x_train, x_test, y_train, _ = train_test_split(X, y, test_size=0.002, random_state=7)
# Fit a random forest regression model
model = create_sklearn_random_forest_regressor(x_train, y_train)
_, cols = x_train.shape
shape = 1, cols
background = csr_matrix(shape, dtype=x_train.dtype)
# Create tabular explainer
exp = tabular_explainer(model, background)
test_logger.info('Running explain global for test_explain_model_sparse_tree')
policy = SamplingPolicy(allow_eval_sampling=True)
exp.explain_global(x_test, sampling_policy=policy)
def test_verify_pipeline_model_coefficient_explanation(self):
# Validate our explainer against an explainable linear model
X, y = shap.datasets.adult()
x_train, x_test, y_train, y_test = train_test_split(X,
y,
test_size=0.2,
random_state=7)
# Note: in pipeline case, we use KernelExplainer;
# in linear case we use LinearExplainer which is much faster
pipeline = [True, False]
threshold = [0.85, 0.76]
for idx, is_pipeline in enumerate(pipeline):
# Fit a logistic regression classifier
model = create_sklearn_logistic_regressor(x_train,
y_train,
pipeline=is_pipeline)
# Create tabular explainer
exp = TabularExplainer(model,
x_train,
features=list(range(x_train.shape[1])))
test_logger.info(
"Running explain model for test_verify_linear_model_coefficient_explanation"
)
# Validate evaluation sampling
policy = {
ExplainParams.SAMPLING_POLICY:
SamplingPolicy(allow_eval_sampling=True)
}
explanation = exp.explain_global(x_test, **policy)
mean_train = np.mean(x_train.values, axis=0)
# Retrieve the model coefficients
if isinstance(model, Pipeline):
model = model.steps[0][1]
coefficients = model.coef_[0]
# Normalize the coefficients by mean for a rough ground-truth of importance
norm_coeff = mean_train * coefficients
# order coefficients by importance
norm_coeff_imp = np.abs(norm_coeff).argsort()[..., ::-1]
# Calculate the correlation
validate_correlation(explanation.global_importance_rank,
norm_coeff_imp, threshold[idx])
def _explain_model_dnn_common(self, tabular_explainer, model, x_train,
x_test, y_train, features):
# Create local tabular explainer without run history
exp = tabular_explainer(model, x_train.values, features=features)
policy = SamplingPolicy(allow_eval_sampling=True)
exp.explain_global(x_test.values, sampling_policy=policy)
def _explain_model_dnn_common(self, tabular_explainer, model, x_train, x_test, y_train, features):
# Create tabular explainer
exp = tabular_explainer(model, x_train.values, features=features, model_task=ModelTask.Classification)
policy = SamplingPolicy(allow_eval_sampling=True)
exp.explain_global(x_test.values, sampling_policy=policy)