retailhero-uplift / uplift_solution.py
dllllb's picture
Upload 9 files
e4de443 verified
raw
history blame
4.23 kB
import pandas
import datetime
import numpy as np
from sklearn.base import clone
from sklearn.model_selection import train_test_split
from sklearn.ensemble import GradientBoostingClassifier
def uplift_fit_predict(model, X_train, treatment_train, target_train, X_test):
"""
Реализация простого способа построения uplift-модели.
Обучаем два бинарных классификатора, которые оценивают вероятность target для клиента:
1. с которым была произведена коммуникация (treatment=1)
2. с которым не было коммуникации (treatment=0)
В качестве оценки uplift для нового клиента берется разница оценок вероятностей:
Predicted Uplift = P(target|treatment=1) - P(target|treatment=0)
"""
X_treatment, y_treatment = X_train[treatment_train == 1, :], target_train[treatment_train == 1]
X_control, y_control = X_train[treatment_train == 0, :], target_train[treatment_train == 0]
model_treatment = clone(model).fit(X_treatment, y_treatment)
model_control = clone(model).fit(X_control, y_control)
predict_treatment = model_treatment.predict_proba(X_test)[:, 1]
predict_control = model_control.predict_proba(X_test)[:, 1]
predict_uplift = predict_treatment - predict_control
return predict_uplift
def uplift_score(prediction, treatment, target, rate=0.3):
"""
Подсчет Uplift Score
"""
order = np.argsort(-prediction)
treatment_n = int((treatment == 1).sum() * rate)
treatment_p = target[order][treatment[order] == 1][:treatment_n].mean()
control_n = int((treatment == 0).sum() * rate)
control_p = target[order][treatment[order] == 0][:control_n].mean()
score = treatment_p - control_p
return score
# Чтение данных
df_clients = pandas.read_csv('data/clients.csv', index_col='client_id')
df_train = pandas.read_csv('data/uplift_train.csv', index_col='client_id')
df_test = pandas.read_csv('data/uplift_test.csv', index_col='client_id')
# Извлечение признаков
df_clients['first_issue_unixtime'] = pandas.to_datetime(df_clients['first_issue_date']).astype(int)/10**9
df_clients['first_redeem_unixtime'] = pandas.to_datetime(df_clients['first_redeem_date']).astype(int)/10**9
df_features = pandas.DataFrame({
'gender_M': (df_clients['gender'] == 'M').astype(int),
'gender_F': (df_clients['gender'] == 'F').astype(int),
'gender_U': (df_clients['gender'] == 'U').astype(int),
'age': df_clients['age'],
'first_issue_time': df_clients['first_issue_unixtime'],
'first_redeem_time': df_clients['first_redeem_unixtime'],
'issue_redeem_delay': df_clients['first_redeem_unixtime'] - df_clients['first_issue_unixtime'],
}).fillna(0)
# Оценка качества на валидации
indices_train = df_train.index
indices_test = df_test.index
indices_learn, indices_valid = train_test_split(df_train.index, test_size=0.3, random_state=123)
valid_uplift = uplift_fit_predict(
model=GradientBoostingClassifier(),
X_train=df_features.loc[indices_learn, :].fillna(0).values,
treatment_train=df_train.loc[indices_learn, 'treatment_flg'].values,
target_train=df_train.loc[indices_learn, 'target'].values,
X_test=df_features.loc[indices_valid, :].fillna(0).values,
)
valid_score = uplift_score(
valid_uplift,
treatment=df_train.loc[indices_valid, 'treatment_flg'].values,
target=df_train.loc[indices_valid, 'target'].values,
)
print('Validation score:', valid_score)
# Подготовка предсказаний для тестовых клиентов
test_uplift = uplift_fit_predict(
model=GradientBoostingClassifier(),
X_train=df_features.loc[indices_train, :].fillna(0).values,
treatment_train=df_train.loc[indices_train, 'treatment_flg'].values,
target_train=df_train.loc[indices_train, 'target'].values,
X_test=df_features.loc[indices_test, :].fillna(0).values,
)
df_submission = pandas.DataFrame({'uplift': test_uplift}, index=df_test.index)
df_submission.to_csv('submission.csv')