def gen_user_hour_click_count(update=False):
'''
生成所有数据的每天没小时点击统计量
拼接键['ID_name', 'clickDay', 'clickHour']
:param update:
:return:
'''
train = load_pickle(raw_data_path + 'train.pkl')
test = load_pickle(raw_data_path + 'test.pkl')
data = train.append(test)
data = addTime(data)
data = addAd(data)
data = addPosition(data)
data = addAppCategories(data)
ads_feature = [
'advertiserID', 'camgaignID', 'adID', 'creativeID', 'appID',
'appCategory'
]
context_feature = ['positionID', 'sitesetID']
state_feature = ads_feature + context_feature
for feature in tqdm(state_feature):
feature_path = feature_data_path + 'user_' + feature + '_click_hour.pkl'
if os.path.exists(feature_path):
print('Found' + feature_path)
else:
print('Generation' + feature_path)
user_feature_click_day = data.groupby(
['userID', 'clickDay', 'clickHour',
feature]).size().reset_index().rename(
columns={0: 'user_' + feature + '_click_hour'})
dump_pickle(user_feature_click_day, feature_path)
def gen_positionID_cvr_smooth(test_day):
"""
* 新学trick：对positionID的在个别时间窗口内的ctr进行贝叶斯平滑处理
"""
feature_path = feature_data_path + 'positionID_cvr_smooth_day_' + str(
test_day) + '.pkl'
if os.path.exists(feature_path):
print('found ' + feature_path)
else:
print('generating ' + feature_path)
data = load_pickle(raw_data_path + 'train.pkl')
data = addTime(data)
positionID_cvr = data[data.clickDay < test_day]
I = positionID_cvr.groupby('positionID')['label'].size().reset_index(
) # size()的作用等于count()，即做计数计算
I.columns = ['positionID', 'I']
C = positionID_cvr.groupby('positionID')['label'].sum().reset_index(
) # 因为点击的label是1，这里做sum()等价于是求出所有的点击量
C.columns = ['positionID', 'C']
positionID_cvr = pd.concat([I, C['C']], axis=1)
hyper = BayesianSmoothing(1, 1)
hyper.update(positionID_cvr['I'].values, positionID_cvr['C'].values,
10000, 0.00000001) # 传入的是展示量和点击量，以及训练的参数设定
alpha = hyper.alpha
beta = hyper.beta
# 根据训练出来的alpha和beta，进行贝叶斯平滑
positionID_cvr['positionID_cvr_smooth'] = (
positionID_cvr['C'] + alpha) / (positionID_cvr['I'] + alpha + beta)
dump_pickle(positionID_cvr[['positionID', 'positionID_cvr_smooth']],
feature_path)
def gen_tricks(start_day, end_day):
'''
生成trick,first_diff,last_diff，install2click，根据gloabl_index拼接
:param start_day:
:param end_day:
:return:
'''
train_data = load_pickle(raw_data_path + 'train.pkl')
test_data = load_pickle(raw_data_path + 'test.pkl')
actions = load_pickle(raw_data_path + 'user_app_actions.pkl')
data = train_data.append(test_data)
del train_data, test_data
data = addTime(data)
data = addAd(data)
for day in tqdm(np.arange(start_day, end_day + 1)):
feature_path = feature_data_path + 'tricks_day_' + str(day) + '.pkl'
if os.path.exists(feature_path):
print('found ' + feature_path)
else:
print('generating ' + feature_path)
df = data.loc[data.clickDay == day]
df = add_trick(df)
df = add_diff(df)
df = add_install2click(df, day, actions)
dump_pickle(
df[[
'global_index', 'trick', 'first_diff', 'last_diff',
'install2click'
]], feature_path)
def gen_ID_global_sum_count(
last_day=27,
stats_features=['positionID', 'creativeID', 'appID', 'adID',
'userID']):
train = load_pickle(raw_data_path + 'train.pkl')
test = load_pickle(raw_data_path + 'test.pkl')
data = train.append(test)
data = addTime(data)
data = data[data.clickDay <= last_day]
del train, test
gc.collect()
data = addAd(data)
data = addPosition(data)
data = addAppCategories(data)
for feature in tqdm(stats_features):
feature_path = feature_data_path + 'global_count_' + feature + '_lastday' + str(
last_day) + '.pkl'
if os.path.exists(feature_path):
print('found ' + feature_path)
# continue
print('generating ' + feature_path)
"""
数据来自：'train.pkl和test.pkl，表示是关于展示和点击的数据
聚合的算子的size()，表示统计的是展示量
"""
feature_count_sum = pd.DataFrame(
data.groupby(feature).size()).reset_index().rename(
columns={0: feature + '_sum_count'})
dump_pickle(feature_count_sum, feature_path)
def gen_hist_cvr_smooth(start_day, end_day, key, alpha=0.25):
# train_data = pd.read_csv(raw_data_path, 'train.csv')
train_data = load_pickle(raw_data_path + 'train.pkl')
test_data = load_pickle(raw_data_path + 'test.pkl')
data = train_data.append(test_data)
del train_data, test_data
gc.collect()
data = addTime(data)
data = addAd(data)
data = addPosition(data)
ID_hist_cvr = None
for day in tqdm(np.arange(start_day, end_day + 1)):
feature_path = feature_data_path + key + '_histcvr_smooth_day_' + str(
day) + '.pkl'
if os.path.exists(feature_path):
print('found ' + feature_path)
else:
print('generating ' + feature_path)
dfCvr = data[data.clickDay < day]
dfCvr = pd.get_dummies(dfCvr, columns=['label'], prefix='label')
dfCvr = dfCvr.groupby([key], as_index=False).sum()
dfCvr[key + '_cvr'] = (dfCvr['label_1'] + alpha) / (
dfCvr['label_0'] + dfCvr['label_0'] + alpha * 2)
# dfCvr['clickDay'] = day
sub_data = pd.merge(data.loc[data.clickDay == day,
['clickDay', key]],
dfCvr[[key, key + '_cvr']],
'left',
on=[
key,
])
sub_data.drop_duplicates(['clickDay', key], inplace=True)
sub_data.sort_values(['clickDay', key], inplace=True)
dump_pickle(sub_data[['clickDay', key, key + '_cvr']],
feature_path)
def gen_CountVector_ID_user_clicks(ID_name,
last_day=27,
ID_describe_feature_names=[
'age_cut', 'gender', 'education',
'marriageStatus', 'haveBaby'
],
drop_na=False):
'''
生成根据train和test表计算的ID_name计数描述向量
拼接键 [ID_name]
:param ID_name:
:param last_day:
:param gen_CountVector_ID_user_clicks:
:param drop_na:
:return:
'''
train = load_pickle(raw_data_path + 'train.pkl')
test = load_pickle(raw_data_path + 'test.pkl')
data = train.append(test)
data = addTime(data)
data = data[data.clickDay <= last_day]
data = addAd(data)
data = addPosition(data)
data = addAppCategories(data)
data = data[['userID', ID_name]]
user_info = pd.read_csv(raw_data_path + 'user.csv')
user_info['age_cut'] = pd.cut(user_info['age'],
bins=[-1, 0, 18, 25, 35, 45, 55, np.inf],
labels=False)
user_info.loc[user_info.education == 7, 'education'] = 6
user_info['hometown_province'] = user_info['hometown'].apply(
lambda x: x // 100)
user_info['residence_province'] = user_info['residence'].apply(
lambda x: x // 100)
for feature in tqdm(ID_describe_feature_names):
feature_path = feature_data_path + 'CountVector_' + ID_name + '_user_clicks_' + feature + '_lastday' + str(
last_day) + '.pkl'
if drop_na:
feature_path += '.no_na'
if os.path.exists(feature_path):
print('found ' + feature_path)
continue
print('generating ' + feature_path)
prefix_name = ID_name + '_user_clicks_' + feature
sub_user_info = pd.get_dummies(user_info[['userID', feature]],
columns=[feature],
prefix=prefix_name)
if drop_na:
sub_user_info.drop([prefix_name + '_0'], axis=1, inplace=True)
data = pd.merge(data, sub_user_info, 'left', 'userID')
dummy_features = sub_user_info.columns.tolist()
dummy_features.remove('userID')
ID_describe_feature = data[[ID_name] + dummy_features].groupby(
[ID_name], as_index=False).sum()
data.drop(dummy_features, axis=1, inplace=True)
dump_pickle(ID_describe_feature, feature_path)
def generate_stats_feature():
'''
输入train和test，进行concat后，添加用户点击数据的统计特征
:return:
'''
feature_path = feature_data_path + 'UserClickStats.pkl'
if os.path.exists(feature_path):
print('Found', feature_path)
else:
train = load_pickle(raw_data_path + 'train.pkl')
test = load_pickle(raw_data_path + 'test.pkl')
data = train.append(test)
del train, test
gc.collect()
data = addTime(data)
data = addAd(data)
data = addPosition(data)
data = addAppCategories(data)
data = add_user_day_click(data)
data = add_user_day_click_count(
data, feature_list=['camgaignID', 'adID', 'appID', 'sitesetID'])
# data = add_user_day_hour_count(data)
# train_origin_features = train.columns.values.tolist()
# test_origin_features = test.columns.values.tolist()
feature_names = [
'user_adID_click_day_mean', # 有些统计特征没包括进来
'user_adID_click_day_min',
'user_camgaignID_click_day_min',
'user_appID_click_day_mean',
'user_appID_click_day_max',
'user_appID_click_day_min',
'user_sitesetID_click_day_mean',
'user_sitesetID_click_day_max',
'user_sitesetID_click_day_min',
'user_click_day_mean',
'user_click_day_max',
'user_click_day_min'
]
print('Generating', feature_path)
columns_day = [
'user_adID_click_day', 'user_camgaignID_click_day',
'user_appID_click_day', 'user_sitesetID_click_day',
'user_click_day'
]
columns_hour = [
'user_adID_click_hour', 'user_camgaignID_click_hour',
'user_appID_click_hour', 'user_sitesetID_click_hour'
]
sub_feature = ['userID', 'clickTime']
# data = pd.concat([train[sub_feature+columns_day+columns_hour],test[sub_feature+columns_day+columns_hour]])
for col in tqdm(columns_day):
data = gen_click_stats(data, col)
# for col in tqdm(columns_day):
# data = add
data = data[feature_names + ['userID']].drop_duplicates(['userID'])
dump_pickle(data, feature_path)
def gen_user_day_click():
feature_path = feature_data_path + 'user_day_clicks.pkl'
if os.path.exists(feature_path):
print('Found' + feature_path)
else:
print('Generating' + feature_path)
train = load_pickle(raw_data_path + 'train.pkl')
test = load_pickle(raw_data_path + 'test.pkl')
all_data = train.append(test)
all_data = addTime(all_data)
user_click_day = pd.DataFrame(
all_data.groupby([
'clickDay', 'userID'
]).size()).reset_index().rename(columns={0: 'user_click_day'})
dump_pickle(user_click_day, feature_path)
def generate_click_trick():
# df['origin_index'] = df.index
feature_path = feature_data_path + 'global_tricks.pkl'
if os.path.exists(feature_path):
print('found ' + feature_path)
else:
train = pd.read_pickle(raw_data_path + 'train.pkl')
test = pd.read_pickle(raw_data_path + 'test.pkl')
df = train.append(test)
df = df[[
'global_index',
'creativeID',
'userID',
'label',
'clickTime',
]]
del train, test
df = addTime(df)
gc.collect()
uct_cnt = df.groupby(['userID', 'creativeID']).size().reset_index()
uct_cnt.rename(columns={0: 'global_uct_cnt'}, inplace=True)
df = pd.merge(df, uct_cnt, how='left', on=['userID', 'creativeID'])
df_1 = df.sort_values(by=['userID', 'clickTime'], ascending=True)
first = df_1.drop_duplicates('userID')
first['global_first'] = 1
first = first[['userID', 'clickTime', 'global_first']]
df = pd.merge(df, first, how='left', on=['userID', 'clickTime'])
df_2 = df.sort_values(by=['userID', 'clickTime'], ascending=False)
last = df_2.drop_duplicates('userID')
last['global_last'] = 1
last = last[['userID', 'clickTime', 'global_last']]
df = pd.merge(df, last, how='left', on=['userID', 'clickTime'])
pd.to_pickle(
df[[
'clickDay',
'global_uct_cnt',
'global_first',
'global_last',
]], feature_path)
def gen_user_day_click():
feature_path = feature_data_path + 'user_day_clicks.pkl'
if os.path.exists(feature_path):
print('Found' + feature_path)
else:
print('Generating' + feature_path)
train = load_pickle(raw_data_path + 'train.pkl')
test = load_pickle(raw_data_path + 'test.pkl')
all_data = train.append(test)
all_data = addTime(all_data) # 添加一些时间维度
"""
.size(): 类似于count(), 统计个数
.reset_index().rename(columns={0:'user_click_day'}): 目的只是为了重命名列名，将聚合统计的结果命名为'user_click_day'
"""
user_click_day = pd.DataFrame(
all_data.groupby([
'clickDay', 'userID'
]).size()).reset_index().rename(columns={0: 'user_click_day'})
# pickle的过程类似于spark的持久化
dump_pickle(user_click_day, feature_path)
def gen_hist_cvr_smooth(start_day, end_day, key, alpha=0.25):
"""
* 新学trick：对传入的时间窗口的每一天的ctr做Laplace Smoothing
"""
train_data = load_pickle(raw_data_path + 'train.pkl')
test_data = load_pickle(raw_data_path + 'test.pkl')
data = train_data.append(test_data)
del train_data, test_data
gc.collect()
data = addTime(data)
data = addAd(data)
data = addPosition(data)
ID_hist_cvr = None
for day in tqdm(np.arange(start_day, end_day + 1)):
feature_path = feature_data_path + key + '_histcvr_smooth_day_' + str(
day) + '.pkl'
if os.path.exists(feature_path):
print('found ' + feature_path)
else:
print('generating ' + feature_path)
dfCvr = data[data.clickDay < day]
dfCvr = pd.get_dummies(dfCvr, columns=['label'], prefix='label')
dfCvr = dfCvr.groupby([key], as_index=False).sum()
"""
- 这里做的是Laplace Smoothing
- 参见https://blog.csdn.net/bbbeoy/article/details/71249316
"""
dfCvr[key + '_cvr'] = (dfCvr['label_1'] + alpha) / (
dfCvr['label_1'] + dfCvr['label_0'] + alpha * 2)
sub_data = pd.merge(data.loc[data.clickDay == day,
['clickDay', key]],
dfCvr[[key, key + '_cvr']],
'left',
on=[
key,
])
sub_data.drop_duplicates(['clickDay', key], inplace=True)
sub_data.sort_values(['clickDay', key], inplace=True)
dump_pickle(sub_data[['clickDay', key, key + '_cvr']],
feature_path)
def gen_positionID_cvr_smooth(test_day):
feature_path = feature_data_path + 'positionID_cvr_smooth_day_' + str(
test_day) + '.pkl'
if os.path.exists(feature_path):
print('found ' + feature_path)
else:
print('generating ' + feature_path)
data = load_pickle(raw_data_path + 'train.pkl')
data = addTime(data)
positionID_cvr = data[data.clickDay < test_day]
I = positionID_cvr.groupby('positionID')['label'].size().reset_index()
I.columns = ['positionID', 'I']
C = positionID_cvr.groupby('positionID')['label'].sum().reset_index()
C.columns = ['positionID', 'C']
positionID_cvr = pd.concat([I, C['C']], axis=1)
hyper = BayesianSmoothing(1, 1)
hyper.update(positionID_cvr['I'].values, positionID_cvr['C'].values,
10000, 0.00000001)
alpha = hyper.alpha
beta = hyper.beta
positionID_cvr['positionID_cvr_smooth'] = (
positionID_cvr['C'] + alpha) / (positionID_cvr['I'] + alpha + beta)
dump_pickle(positionID_cvr[['positionID', 'positionID_cvr_smooth']],
feature_path)
