def run_cochran_q_test(y_test, *model_predictions, output_name):
"""
Runs Cochran's Q test to determine if there is a statistically significant difference in more than two models' class
predictions. The function can support up to five sets of predictions. Results are saved locally.
:param y_test: y_test series
:param model_predictions: arbitrary number of model predictions
:param output_name: name to append to file to identify models used in the test
"""
n_models = len(model_predictions)
if n_models == 3:
chi2, p = cochrans_q(y_test.values, model_predictions[0].values,
model_predictions[1].values,
model_predictions[2].values)
elif n_models == 4:
chi2, p = cochrans_q(y_test.values, model_predictions[0].values,
model_predictions[1].values,
model_predictions[2].values,
model_predictions[3].values)
elif n_models == 5:
chi2, p = cochrans_q(y_test.values, model_predictions[0].values,
model_predictions[1].values,
model_predictions[2].values,
model_predictions[3].values,
model_predictions[4].values)
else:
raise Exception(
'function cannot support more than five sets of predictions')
pd.DataFrame({
'chi2': [chi2],
'p': [p]
}).to_csv(
os.path.join('modeling', 'comparison_files',
f'{output_name}_cochrans_q_test.csv'))
def test_compare_to_mcnemar_on_2_models():
y_true = np.array([0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0])
ym1 = np.array([1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0])
ym2 = np.array([1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0])
q, p = cochrans_q(y_true, ym1, ym2)
mcn_q, mcn_p = mcnemar(mcnemar_table(y_true, ym1, ym2),
corrected=False,
exact=False)
assert q == mcn_q
assert p == mcn_p
def test_on_dataset():
y_true = np.array([0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0])
ym1 = np.array([1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0])
ym2 = np.array([1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0])
ym3 = np.array([1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1, 1])
q, p_value = cochrans_q(y_true, ym1, ym2, ym3)
assert round(q, 3) == 7.529
assert round(p_value, 3) == 0.023
def test_compare_to_mcnemar_on_2_models():
y_true = np.array([0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0])
ym1 = np.array([1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0])
ym2 = np.array([1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0])
q, p = cochrans_q(y_true, ym1, ym2)
mcn_q, mcn_p = mcnemar(mcnemar_table(y_true, ym1, ym2),
corrected=False,
exact=False)
assert q == mcn_q
assert p == mcn_p
def test_on_dataset():
y_true = np.array([0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0])
ym1 = np.array([1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0])
ym2 = np.array([1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0])
ym3 = np.array([1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1, 1])
q, p_value = cochrans_q(y_true, ym1, ym2, ym3)
assert round(q, 3) == 7.529
assert round(p_value, 3) == 0.023
def summarize_feature_comparisons(
base_clf: BaseEstimator, comparison_clfs: Dict[str, BaseEstimator], X_test, y_test
):
from mlxtend.evaluate import mcnemar, cochrans_q, mcnemar_table
summary_dict = collections.OrderedDict()
mcnemar_tbs = dict()
# create list of predicted values
base_y_predict = base_clf.predict(X_test)
y_predictions = [base_y_predict]
for idx, (name, clf) in enumerate(comparison_clfs.items()):
# get the probability
y_predict_proba = clf.predict_proba(X_test)
y_predict = clf.predict(X_test)
# form mcnemar tables against base classifier
tb = mcnemar_table(y_test, base_y_predict, y_predict)
mcnemar_tbs[f"base vs {name}"] = tb.values()
# store predictions per classifier
y_predictions.append(y_predict)
# first run cochrans Q test
qstat, pval = cochrans_q(y_test, *y_predictions)
summary_dict["cochrans_q"] = qstat
summary_dict["cochrans_q_pval"] = pval
# run mcnemars test against all the predictions
for name, table in mcnemar_tbs.items():
chi2stat, pval = mcnemar(table, exact=True)
summary_dict[f"mcnemar_{name}_chi2stat"] = chi2stat
summary_dict[f"mcnemar_{name}_pval"] = pval
return summary_dict
# Implementing GridSearchCV on Logistic Regression model
model2 = GridSearchCV(lr, parameters_lr, cv = cv)
model2.fit(X_train, y_train) # Fitting on training data
# In[16]:
# Implementing GridSearchCV on MLP classifier
model3 = GridSearchCV(mlp, parameters_mlp, cv = cv)
model3.fit(X_train, y_train) # Fitting on training data
# In[17]:
y_model1 = model1.predict(X_test)
y_model2 = model2.predict(X_test)
y_model3 = model3.predict(X_test)
y_test = np.array(y_test)
# In[18]:
q, p_value = cochrans_q(y_test, y_model1, y_model2, y_model3)
print('Q: %.3f' % q)
print('p-value: %.3f' % p_value)
result = f_oneway(df_result['uni_R'].to_numpy(),
df_result['bi_R'].to_numpy(),
df_result['unibi_R'].to_numpy())
print("ANNOVA R : %0.5f, %0.5f" % result)
result = f_oneway(df_result['uni_F1'].to_numpy(),
df_result['bi_F1'].to_numpy(),
df_result['unibi_F1'].to_numpy())
print("ANNOVA F1 : %0.5f, %0.5f" % result)
# Coher Q analysis
y_uni = sr_uni.to_numpy()
y_bi = sr_bi.to_numpy()
y_unibi = sr_unibi.to_numpy()
q, p_value = cochrans_q(y, y_uni, y_bi, y_unibi)
print("COHRAN Q-Test: q: %0.5f, p_value: %0.5f" % (q, p_value))
l_grams = ['uni', 'bi', 'unibi']
l_rslt = [y_uni, y_bi, y_unibi]
l_pair = list(zip(l_grams, l_rslt))
l_mcnemar_rslt = []
for i, t0 in enumerate(l_pair):
for j, t1 in enumerate(l_pair[i + 1:]):
k0 = t0[0]
k1 = t1[0]
v0 = t0[1]
v1 = t1[1]
tb = mcnemar_table(y_target=y, y_model1=v0, y_model2=v1)
print("ANNOVA R : %0.5f, %0.5f" % result)
result = f_oneway(
df_result['tc_lower_F1'].to_numpy(),
df_result['tc_swrem_F1'].to_numpy(),
df_result['tc_stem_F1'].to_numpy(),
df_result['tc_swrem_stem_F1'].to_numpy(),
)
print("ANNOVA F1 : %0.5f, %0.5f" % result)
# COHRAN Q-Test
y_tc_lower = sr_tc_lower.to_numpy()
y_tc_swrem = sr_tc_swrem.to_numpy()
y_tc_stem = sr_tc_stem.to_numpy()
y_tc_swrem_stem = sr_tc_swrem_stem.to_numpy()
q, p_value = cochrans_q(y, y_tc_lower, y_tc_swrem, y_tc_stem,
y_tc_swrem_stem)
print("COHRAN Q-Test: q: %0.5f, p_value: %0.5f" % (q, p_value))
l_repr = ['tc_lower', 'tc_swrem', 'tc_stem', 'tc_swrem_stem']
l_rslt = [y_tc_lower, y_tc_swrem, y_tc_stem, y_tc_swrem_stem]
l_pair = list(zip(l_repr, l_rslt))
l_mcnemar_rslt = []
for i, t0 in enumerate(l_pair):
for j, t1 in enumerate(l_pair[i + 1:]):
k0 = t0[0]
k1 = t1[0]
v0 = t0[1]
v1 = t1[1]
tb = mcnemar_table(y_target=y, y_model1=v0, y_model2=v1)
print("ANNOVA R : %0.5f, %0.5f" % result)
result = f_oneway(df_result['multi_nb_F1'].to_numpy(),
df_result['svc_F1'].to_numpy(),
df_result['lsvc_F1'].to_numpy(),
df_result['rf_F1'].to_numpy(),
df_result['lr_F1'].to_numpy(),
df_result['ada_F1'].to_numpy())
print("ANNOVA F1 : %0.5f, %0.5f" % result)
# Coher Q analysis
q, p_value = cochrans_q(
y,
sr_multi_nb.to_numpy(),
sr_svc.to_numpy(),
sr_lsvc.to_numpy(),
sr_rf.to_numpy(),
sr_lr.to_numpy(),
sr_ada.to_numpy(),
)
print("COHRAN Q-Test: q: %0.5f, p_value: %0.5f" % (q, p_value))
f_out = basename(f_in)
df_result.to_excel(f_in.replace('.json', '_RESULT.xlsx'))
l_algo = ['multi_nb', 'svc', 'lsvc', 'rf', 'lr', 'ada']
l_rslt = [
sr_multi_nb.to_numpy(),
sr_svc.to_numpy(),
sr_lsvc.to_numpy(),
print("ANNOVA R : %0.5f, %0.5f" % result)
str_annova_f += "%0.5f " % (result[0])
str_annova_p += "%0.5f " % (result[1])
result = f_oneway(df_result['t_lower_F1'].to_numpy(),
df_result['c_lower_F1'].to_numpy(),
df_result['tc_lower_F1'].to_numpy())
print("ANNOVA F1 : %0.5f, %0.5f" % result)
str_annova_f += "%0.5f " % (result[0])
str_annova_p += "%0.5f " % (result[1])
# COHRAN Q-Test
y_tc_lower = sr_tc_lower.to_numpy()
y_t_lower = sr_t_lower.to_numpy()
y_c_lower = sr_c_lower.to_numpy()
q, p_value = cochrans_q(y, y_tc_lower, y_t_lower, y_c_lower)
print("COHRAN Q-Test: q: %0.5f, p_value: %0.5f" % (q, p_value))
l_repr = ['t_lower', 'c_lower', 'tc_lower']
l_rslt = [y_t_lower, y_c_lower, y_tc_lower]
l_pair = list(zip(l_repr, l_rslt))
str_result = "MCNEMAR RESULT: "
for i, t0 in enumerate(l_pair):
for j, t1 in enumerate(l_pair[i + 1:]):
k0 = t0[0]
k1 = t1[0]
v0 = t0[1]
v1 = t1[1]
