def test_enable_ipw_without_known_propensity_parallel_runner():
seed = 0
df = generate_data(
N=1000,
n_features=3,
beta=[0, -2, 3, -5], # Effect of [intercept and features] on outcome
error_std=0.1,
tau=[1, -5, -5, 10], # Effect of [intercept and features] on treated outcome
tau_std=0.1,
discrete_outcome=True,
seed=seed,
feature_effect=0, # Effect of beta on treated outxome
propensity_coef=[
0,
-1,
1,
-1,
], # Effect of [intercept and features] on propensity log-odds for treatment
index_name="index",
)
train_df, test_df = train_test_split(df, test_size=0.2, stratify=df["Treatment"])
cl = CausalLift(
train_df, test_df, enable_ipw=True, verbose=3, runner="ParallelRunner"
)
train_df, test_df = cl.estimate_cate_by_2_models()
estimated_effect_df = cl.estimate_recommendation_impact()
assert isinstance(train_df, pd.DataFrame)
assert isinstance(test_df, pd.DataFrame)
assert isinstance(estimated_effect_df, pd.DataFrame)
def test_enable_ipw_with_known_propensity():
seed = 0
df = generate_data(
N=1000,
n_features=3,
beta=[0, -2, 3, -5], # Effect of [intercept and features] on outcome
error_std=0.1,
tau=[1, -5, -5,
10], # Effect of [intercept and features] on treated outcome
tau_std=0.1,
discrete_outcome=True,
seed=seed,
feature_effect=0, # Effect of beta on treated outxome
propensity_coef=[
0,
-1,
1,
-1,
], # Effect of [intercept and features] on propensity log-odds for treatment
index_name="index",
)
train_df, test_df = train_test_split(df,
test_size=0.2,
random_state=seed,
stratify=df["Treatment"])
test_random_propensity = True
if test_random_propensity:
import random
train_df = train_df.copy()
train_df.loc[:, "Propensity"] = [
random.random() for _ in range(train_df.shape[0])
]
test_df = test_df.copy()
test_df.loc[:, "Propensity"] = [
random.random() for _ in range(test_df.shape[0])
]
cl = CausalLift(train_df, test_df, enable_ipw=True, verbose=3)
train_df, test_df = cl.estimate_cate_by_2_models()
estimated_effect_df = cl.estimate_recommendation_impact()
assert isinstance(train_df, pd.DataFrame)
assert isinstance(test_df, pd.DataFrame)
assert isinstance(estimated_effect_df, pd.DataFrame)
from causallift import CausalLift
from sklearn.model_selection import train_test_split
seed = 0
from causallift import generate_data
df = generate_data( \
N=1000,
n_features=3,
beta=[0, -2, 3, -5], # Effect of [intercept and features] on outcome
error_std=0.1,
tau=[1, -5, -5, 10], # Effect of [intercept and features] on treated outcome
tau_std=0.1,
discrete_outcome=True,
seed=seed,
feature_effect=0, # Effect of beta on treated outxome
propensity_coef=[0, -1, 1, -1], # Effect of [intercept and features] on propensity log-odds for treatment
index_name='index')
train_df, test_df = train_test_split(df,
test_size=0.2,
random_state=seed,
stratify=df['Treatment'])
cl = CausalLift(train_df, test_df, enable_ipw=False, random_state=0, verbose=3)
train_df, test_df = cl.estimate_cate_by_2_models()
estimated_effect_df = cl.estimate_recommendation_impact()
def test_skopt():
seed = 0
df = generate_data(
N=1000,
n_features=3,
beta=[0, -2, 3, -5], # Effect of [intercept and features] on outcome
error_std=0.1,
tau=[1, -5, -5,
10], # Effect of [intercept and features] on treated outcome
tau_std=0.1,
discrete_outcome=True,
seed=seed,
feature_effect=0, # Effect of beta on treated outxome
propensity_coef=[
0,
-1,
1,
-1,
], # Effect of [intercept and features] on propensity log-odds for treatment
index_name="index",
)
train_df, test_df = train_test_split(df,
test_size=0.2,
random_state=seed,
stratify=df["Treatment"])
uplift_model_params = dict(
search_cv="skopt.BayesSearchCV",
estimator="xgboost.XGBClassifier",
const_params=dict(
booster="gbtree",
silent=True,
objective="binary:logistic",
base_score=0.5,
eval_metric="auc",
n_jobs=-1,
seed=seed,
),
search_spaces=dict(n_estimators=Integer(100, 200)),
random_state=seed,
scoring="roc_auc",
cv=3,
n_jobs=-1,
n_iter=10,
verbose=1,
refit=True,
)
cl = CausalLift(
train_df,
test_df,
enable_ipw=True,
verbose=3,
uplift_model_params=uplift_model_params,
)
train_df, test_df = cl.estimate_cate_by_2_models()
estimated_effect_df = cl.estimate_recommendation_impact()
assert isinstance(train_df, pd.DataFrame)
assert isinstance(test_df, pd.DataFrame)
assert isinstance(estimated_effect_df, pd.DataFrame)
seed = 0
df = generate_data(
N=1000,
n_features=3,
beta=[0, -2, 3, -5], # Effect of [intercept and features] on outcome
error_std=0.1,
tau=[1, -5, -5,
10], # Effect of [intercept and features] on treated outcome
tau_std=0.1,
discrete_outcome=True,
seed=seed,
feature_effect=0, # Effect of beta on treated outxome
propensity_coef=[
0,
-1,
1,
-1,
], # Effect of [intercept and features] on propensity log-odds for treatment
index_name="index",
)
train_df, test_df = train_test_split(df,
test_size=0.2,
random_state=seed,
stratify=df["Treatment"])
estimator = XGBClassifier(
booster="gbtree",
silent=True,
objective="binary:logistic",
base_score=0.5,
eval_metric="auc",
n_jobs=-1,
seed=seed,
)
model = BayesSearchCV(
estimator=estimator,
search_spaces=dict(n_estimators=Integer(100, 200)),
random_state=seed,
scoring="roc_auc",
cv=3,
n_jobs=-1,
n_iter=10,
verbose=1,
refit=True,
)
cl_ext = CausalLift(train_df,
test_df,
enable_ipw=True,
verbose=3,
uplift_model_params=model)
train_df_ext, test_df_ext = cl_ext.estimate_cate_by_2_models()
estimated_effect_df_ext = cl_ext.estimate_recommendation_impact()
assert isinstance(train_df_ext, pd.DataFrame)
assert isinstance(test_df_ext, pd.DataFrame)
assert isinstance(estimated_effect_df_ext, pd.DataFrame)
def causal_lift_model(cm_df, decoder, treatment_name, use_multi=True, seed=5):
# Need to drop null for computations - could also impute these
original_len = len(cm_df)
print('Original data length {}'.format(original_len))
cm_df[treatment_name] = cm_df[treatment_name].dropna().reset_index(
drop=True)
clean_len = len(cm_df)
print('Clean data length {}'.format(clean_len))
ratio_dropped = (original_len - clean_len) / clean_len
# Separate treatment into two categories (below or above average)
treatment_below_avg = cm_df[treatment_name].mean()
cm_df['Treatment'] = np.where(cm_df[treatment_name] <= treatment_below_avg,
1, 0)
# Use multiple classes as target
if use_multi:
cm_df['Outcome'] = cm_df['Description']
# Use the top 10 features
cm_df = cm_df[[
'Treatment', 'Outcome', 'Weapon_FIREARM', 'Weapon_HANDS',
'Weapon_KNIFE', 'Weapon_NONE', 'Weapon_OTHER', 'Neighborhood',
'Premise', 'Month', 'Hour', 'Outside'
]]
train_df, test_df = train_test_split(cm_df,
test_size=0.2,
random_state=seed,
stratify=cm_df['Treatment'])
print(
'\n[Estimate propensity scores for Inverse Probability Weighting.]'
)
CausalLift(train_df, test_df, enable_ipw=True, verbose=3)
# Use each class as a binary outcome
else:
# Get each outcome and loop through them testing them as binary variables
outcome_names = set(cm_df['Description'].values)
print(outcome_names)
for outcome in outcome_names:
cm_df_copy = cm_df.copy()
print('\n{}'.format(decoder[outcome]))
cm_df_copy['Outcome'] = np.where(
cm_df_copy['Description'] == outcome, 1, 0)
print(cm_df_copy['Outcome'])
# Use the top 10 features
cm_df_copy = cm_df_copy[[
'Treatment', 'Outcome', 'Weapon_FIREARM', 'Weapon_HANDS',
'Weapon_KNIFE', 'Weapon_NONE', 'Weapon_OTHER', 'Neighborhood',
'Premise', 'Month', 'Hour', 'Outside'
]]
train_df, test_df = train_test_split(
cm_df_copy,
test_size=0.2,
random_state=seed,
stratify=cm_df_copy['Treatment'])
print(
'\n[Estimate propensity scores for Inverse Probability Weighting.]'
)
CausalLift(train_df, test_df, enable_ipw=True, verbose=3)