def test_generate_local_sample(self):
row = self.frame.iloc[0, :]
explainer = LIMEExplainer(training_frame=self.frame,
X=self.X,
model=self.model)
local_sample = explainer._generate_local_sample(row)
N = explainer.N
self.assertAlmostEqual(local_sample.loc[:, 'MSSubClass'].std(),
self.frame.loc[:, 'MSSubClass'].std(),
places=0)
self.assertEqual(local_sample.shape[0], N)
del explainer
def test_explain_w_o_discretize(self):
row_id = 0
explainer = LIMEExplainer(training_frame=self.frame,
X=self.X,
model=self.model,
discretize=None)
explainer.explain(row_id)
self.assertAlmostEqual(
explainer.lime_pred,
explainer.lime.coef()['Intercept'] +
explainer.reason_code_values['Local Contribution'].sum())
del explainer
def test_local_contrib_w_discretize(self):
row = self.frame.iloc[0, :]
explainer = LIMEExplainer(training_frame=self.frame,
X=self.X,
model=self.model,
discretize=self.discretize)
local_sample = explainer._generate_local_sample(row)
scored_local_sample = \
explainer._score_local_sample(local_sample,
row[local_sample.columns])
weighted_scored_local_sample = \
explainer._calculate_distance_weights(0,
scored_local_sample)
discretized_weighted_scored_local_sample = \
explainer._discretize_numeric(weighted_scored_local_sample)
disc_row = pd.DataFrame(columns=self.X)
for name in self.discretize:
disc_row[name] = pd.cut(pd.Series(row[name]),
bins=explainer.bins_dict[name])
not_in = list(set(self.X) - set(self.discretize))
disc_row[not_in] = row[not_in].values
explainer.lime = \
explainer._regress(discretized_weighted_scored_local_sample,
h2o.H2OFrame(disc_row))
self.assertEqual(explainer.reason_code_values.shape, (21, 2))
del explainer
def test_score_local_sample(self):
row = self.frame.iloc[0, :]
explainer = LIMEExplainer(training_frame=self.frame,
X=self.X,
model=self.model)
local_sample = explainer._generate_local_sample(row)
scored_local_sample = \
explainer._score_local_sample(local_sample,
row[local_sample.columns])
self.assertEqual(scored_local_sample.shape[1],
local_sample.shape[1] + 1)
self.assertEqual(scored_local_sample.columns[-1], 'predict')
del explainer
def test_calculate_distance_weights(self):
row = self.frame.iloc[0, :]
explainer = LIMEExplainer(training_frame=self.frame,
X=self.X,
model=self.model)
local_sample = explainer._generate_local_sample(row)
scored_local_sample = \
explainer._score_local_sample(local_sample,
row[local_sample.columns])
weighted_scored_local_sample = \
explainer._calculate_distance_weights(0,
scored_local_sample)
self.assertEqual(weighted_scored_local_sample.shape[1],
local_sample.shape[1] + 2)
self.assertEqual(weighted_scored_local_sample.columns[-1], 'distance')
del explainer
def test_regress_w_discretize(self):
row = self.frame.iloc[0, :]
explainer = LIMEExplainer(training_frame=self.frame,
X=self.X,
model=self.model,
discretize=self.discretize)
local_sample = explainer._generate_local_sample(row)
scored_local_sample = \
explainer._score_local_sample(local_sample,
row[local_sample.columns])
weighted_scored_local_sample = \
explainer._calculate_distance_weights(0,
scored_local_sample)
discretized_weighted_scored_local_sample = \
explainer._discretize_numeric(weighted_scored_local_sample)
disc_row = pd.DataFrame(columns=self.X)
for name in self.discretize:
disc_row[name] = pd.cut(pd.Series(row[name]),
bins=explainer.bins_dict[name])
not_in = list(set(self.X) - set(self.discretize))
disc_row[not_in] = row[not_in].values
lime = explainer._regress(discretized_weighted_scored_local_sample,
h2o.H2OFrame(disc_row))
self.assertTrue(explainer.discretize)
self.assertIsNotNone(lime)
self.assertAlmostEqual(0.9859272974096093, explainer.lime_r2)
del explainer
def test_local_contrib_w_o_discretize(self):
row = self.frame.iloc[0, :]
explainer = LIMEExplainer(training_frame=self.frame,
X=self.X,
model=self.model)
local_sample = explainer._generate_local_sample(row)
scored_local_sample = \
explainer._score_local_sample(local_sample,
row[local_sample.columns])
weighted_scored_local_sample = \
explainer._calculate_distance_weights(0,
scored_local_sample)
explainer.lime = \
explainer._regress(weighted_scored_local_sample,
h2o.H2OFrame(pd.DataFrame(row).T))
self.assertEqual(explainer.reason_code_values.shape, (26, 2))
del explainer
def test_regress_w_o_discretize(self):
row = self.frame.iloc[0, :]
explainer = LIMEExplainer(training_frame=self.frame,
X=self.X,
model=self.model)
local_sample = explainer._generate_local_sample(row)
scored_local_sample = \
explainer._score_local_sample(local_sample,
row[local_sample.columns])
weighted_scored_local_sample = \
explainer._calculate_distance_weights(0,
scored_local_sample)
lime = explainer._regress(weighted_scored_local_sample,
h2o.H2OFrame(pd.DataFrame(row).T))
self.assertFalse(explainer.discretize)
self.assertIsNotNone(lime)
self.assertAlmostEqual(0.9649889709835218, explainer.lime_r2)
del explainer
def test_discretize_numeric(self):
row = self.frame.iloc[0, :]
explainer = LIMEExplainer(training_frame=self.frame,
X=self.X,
model=self.model,
discretize=self.discretize)
N = explainer.N
local_sample = explainer._generate_local_sample(row)
scored_local_sample = \
explainer._score_local_sample(local_sample,
row[local_sample.columns])
weighted_scored_local_sample = \
explainer._calculate_distance_weights(0,
scored_local_sample)
discretized_weighted_scored_local_sample = \
explainer._discretize_numeric(weighted_scored_local_sample)
self.assertEqual(discretized_weighted_scored_local_sample.shape,
(N, local_sample.shape[1] + 2))
self.assertEqual(
discretized_weighted_scored_local_sample[
self.discretize].dtypes.unique()[0], 'category')
del explainer