-
Notifications
You must be signed in to change notification settings - Fork 25
/
herbert.py
176 lines (150 loc) · 7.89 KB
/
herbert.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
# coding=utf-8
# @author: herbert-chen
# github: https://github.com/Herbert95/JDATA_
import pandas as pd
import numpy as np
import datetime
import copy
import xgboost as xgb
from sklearn.model_selection import train_test_split
from tqdm import tqdm
import warnings
warnings.filterwarnings('ignore')
sku = pd.read_csv('./jdata_sku_basic_info.csv', )
action = pd.read_csv('./jdata_user_action.csv', parse_dates=['a_date'])
basic_info = pd.read_csv('./jdata_user_basic_info.csv')
comment_score = pd.read_csv('./jdata_user_comment_score.csv', parse_dates=['comment_create_tm'])
order = pd.read_csv('./jdata_user_order.csv', parse_dates=['o_date'])
order = pd.merge(order, sku, on='sku_id', how='left')
order = pd.merge(order, basic_info, on='user_id', how='left')
action = pd.merge(action, sku, how='left', on='sku_id')
order['month'] = order['o_date'].apply(lambda x: x.month)
action['month'] = action['a_date'].apply(lambda x: x.month)
# 评分函数
def score(pred, real):
# pred: user_id, pre_date | real: user_id, o_date
# wi与oi的定义与官网相同
pred['pred_day'] = pd.to_datetime(pred['pred_date']).dt.day
pred['index'] = np.arange(pred.shape[0]) + 1
pred['wi'] = 1 / (1 + np.log(pred['index']))
real['real_day'] = pd.to_datetime(real['o_date']).dt.day
real['oi'] = 1
compare = pd.merge(pred, real, how='left', on='user_id')
compare.fillna(0, inplace=True) # 实际上没有购买的用户,correct_for_S1列的值为nan,将其赋为0
S1 = np.sum(compare['oi'] * compare['wi']) / np.sum(compare['wi'])
compare_for_S2 = compare[compare['oi'] == 1]
S2 = np.sum(10 / (10 + np.square(compare_for_S2['pred_day'] - compare_for_S2['real_day']))) / real.shape[0]
S = 0.4 * S1 + 0.6 * S2
print("S1=", S1, "| S2 ", S2)
print("S =", S)
# 三月训练,四月验证;四月训练,五月验证
# def main(evaluate=False):
evaluate = False
train_month = 4 - int(evaluate)
if evaluate:
train_action = action[action['month'] != 3][action['month'] != 4]
else:
train_action = action[action['month'] != 4]
# 构建训练集:是首次购买日期,所以训练集只取训练月份最早购买的那一天
train_data = order[order['month'] == train_month][['user_id', 'o_date', 'cate']].sort_values(by=['user_id', 'o_date']).drop_duplicates()
train_data = train_data.drop(train_data[train_data[['user_id', 'cate']].duplicated()].index, axis=0)
train_data = pd.merge(train_data, basic_info, on='user_id', how='left')
# 增加训练集负样本
print('creating train dataset')
all_negative_train_data = pd.DataFrame()
len_of_original_train_data = train_data.shape[0]
for day_shift in [-2, -1, 1, 2]:
negative_train_data = copy.deepcopy(train_data)
negative_train_data['o_date'] = negative_train_data['o_date'].apply(lambda x: x + datetime.timedelta(days=day_shift))
all_negative_train_data = pd.concat([all_negative_train_data, negative_train_data])
train_data['label'] = 1
all_negative_train_data['label'] = 0
all_negative_train_data['month'] = all_negative_train_data['o_date'].apply(lambda x: x.month)
all_negative_train_data = all_negative_train_data[all_negative_train_data['month'] == train_month]
# 删去错误的负样本(删除条件:若当天有购买,不应设其label为0)
wrong_data_index = pd.concat([train_data[['user_id', 'cate', 'o_date']],
all_negative_train_data[['user_id', 'cate', 'o_date']]]).reset_index(drop=True).duplicated()
wrong_data_index[:len_of_original_train_data] = False
train_data = pd.concat([train_data, all_negative_train_data]).reset_index(drop=True)
train_data = train_data.drop(train_data[wrong_data_index].index, axis=0)
train_data = train_data.sample(frac=1, random_state=777)
train_data['day'] = train_data['o_date'].apply(lambda x: x.day)
# 构建测试集
print('creating test dataset')
test_data_cate_101, test_data_cate_30 = basic_info[['user_id']], basic_info[['user_id']]
test_data_cate_101['cate'] = 101
test_data_cate_30['cate'] = 30
original_test_data = pd.concat([test_data_cate_101, test_data_cate_30], axis=0)
original_test_data = pd.merge(original_test_data, basic_info, on='user_id', how='left')
test_data = pd.DataFrame()
for i in tqdm(range({4:30, 5:31}[train_month+1])):
test_data_per_day = copy.deepcopy(original_test_data)
test_data_per_day['o_date'] = datetime.datetime(2017, train_month + 1, i + 1)
test_data_per_day['day'] = i + 1
test_data = pd.concat([test_data, test_data_per_day])
# 构建action特征
base_id = ['user_id', 'cate', 'o_date']
def add_action_feature(data, action_data, mode):
first_day = {'train': datetime.datetime(2017, train_month, 1), 'test': datetime.datetime(2017, train_month + 1, 1)}[mode]
current_month = {'train': train_month, 'test': train_month + 1}[mode]
pre_data = pd.DataFrame()
for action_index, action in enumerate(['look', 'star']):
# 用户从上次浏览或关注到现在的时间
latest_action_date, days_from_latest_action_date = 'latest_%s_date' % action, 'days_from_latest_%s_date' % action
print('adding %s feature for %s data' % (days_from_latest_action_date, mode))
action_data_i = action_data[action_data['a_type'] == action_index + 1]
action_data_i = action_data_i.groupby(['user_id', 'cate']).a_date.agg({latest_action_date: max}).reset_index()
data = pd.merge(data, action_data_i, how='left', on=['user_id', 'cate'])
data[latest_action_date].fillna(datetime.datetime(2000, 1, 1), inplace=True)
data_for_compute_action_days = data[[latest_action_date]].drop_duplicates()
data_for_compute_action_days[days_from_latest_action_date] = data_for_compute_action_days[
latest_action_date].apply(lambda x: (first_day - x).days)
data = pd.merge(data, data_for_compute_action_days, how='left', on=latest_action_date)
data[days_from_latest_action_date] = data[days_from_latest_action_date] + data['day'] - 1
return data
train_data = add_action_feature(train_data, train_action, 'train')
test_data = add_action_feature(test_data, action, 'test')
# one hot
len_of_total_train_data = train_data.shape[0]
all_data = pd.concat([train_data, test_data])
one_hot_feature = ['cate', 'sex']
all_data = pd.get_dummies(all_data, columns=one_hot_feature)
# 处理缺失数据
all_data['age'] = all_data['age'].replace(-1, all_data['age'].mode()[0])
all_data.fillna(0, inplace=True)
train_data = all_data[:len_of_total_train_data]
test_data = all_data[len_of_total_train_data:]
# 筛选特征
cols_to_delete = ['user_id', 'sku_id', 'o_id', 'o_date', 'o_area', 'o_sku_num', 'month', 'latest_look_date', 'latest_star_date', 'day']
feature_list = [feature for feature in test_data.columns if feature not in cols_to_delete]
# xgboost参数
params = {'booster': 'gbtree',
'objective': 'binary:logistic',
'gamma': 0.025,
'min_child_weight': 6,
'max_depth': 7,
'lambda': 1,
'subsample': 0.7,
'colsample_bytree': 0.6,
'eta': 0.1,
'seed': 0,
'max_delta_step': 0.5,
'silent': 0,
'scale_pos_weight': 4,
}
# 输出
train_xgb = xgb.DMatrix(train_data[feature_list].values, train_data['label'])
test_xgb = xgb.DMatrix(test_data[feature_list].values)
model_xgb = xgb.train(params, train_xgb, num_boost_round=500)
test_y_xgb = model_xgb.predict(test_xgb)
result = test_data[['user_id']]
result['pred_date'] = test_data['o_date']
result['prob'] = test_y_xgb
result = result.sort_values(by=['prob', 'user_id'], ascending=False)
result = result.drop(result[result[['user_id']].duplicated()].index, axis=0)
if evaluate:
real_result = order[order['month'] == train_month + 1][['user_id', 'o_date']].sort_values(by=['user_id', 'o_date']).drop_duplicates()
real_result = real_result.drop(real_result[real_result[['user_id']].duplicated()].index, axis=0)
score(result[:50000], real_result)
else:
result[['user_id', 'pred_date']][:50000].to_csv('./result.csv', index=None)