上一次，我们介绍了电子商务AI算法大赛中我个人的解决方案。虽然在经过一顿操作后，线上分数得到了肉眼可见的提升，但相比于大佬们的成绩还是差了亿点点的
Alt text

所以呢，这次来解读top2选手的解决方案，从优秀开源代码中学习思路，积累经验。

关于赛题的介绍以及数据集说明，可以回看上一篇文章哈：xxxxxx

线下验证方案

date字段代表的是第几天，训练集中date的取值范围是从1 到116，共116天；测试集中date的取值范围是117到123共7天。于是从训练集中划出最后7天的数据作为验证集进行线下验证：

Alt text

而在最终提交时，取消验证集，用全部训练数据进行训练：

Alt text

数据集划分代码如下：

1
2
3
4
5
6
def split_dataset(data, valid_start_date, gap):
    train = data[data['date']<valid_start_date].copy()
    valid = data[(data['date']>=valid_start_date)&(data['date']<valid_start_date+gap)].copy()
    test = data[data['date']>=valid_start_date+gap].copy()
    print(f'训练集样本数: {train.shape[0]}, 验证集样本数: {valid.shape[0]}, 测试集样本数: {test.shape[0]}')
    return train, valid, test

调用时：

1
train, valid, test = split_dataset(data=data, valid_start_date=110, gap=7) # 划分训练\验证集

魔法特征

魔法特征，就是能够给模型带来大幅度涨分的特征。

绘制article_id的折线图：

1
2
3
4
5
6
import matplotlib.pyplot as plt
y=train_df['article_id'][:200]
x=[i for i in range(len(y))]
plt.plot(x,y)
plt.ylabel('article_id')
plt.savefig('article_id.png',dpi=600)

Alt text

观察上图，发现article_id的取值呈现周期性变化。

纵坐标article_id指的是文章id取值，而横坐标是时间推移，这里的时间在一天内。

而如果绘制全部116天的article_id折线图，这种周期性规律就不复存在了：
Alt text

所以，猜想article_id这个特征应该是与时间有关的，具体地，某件商品的销量在一天内会呈现周期性变动。

既然这种规律是存在于一天内的，那自然就想到按照天数进行分组，同一天的数据归为同一组。

然后对每组内的article_id进行排序，就得到了一个新的特征：magic_rank

1
data['magic_rank'] = data.groupby(['date'])['article_id'].rank()

基于该思想，可以构建出许多有用的新特征。

比如，将date改为其它特征，将rank改为常见的统计量

1
2
3
g = data.groupby(f)#f是某个特征，比如`mall`，`author`
for stat in ['max', 'min', 'mean']:
    data[f'magic_{name}_id_{stat}'] = g['article_id'].transform(stat)

再比如，将date改为其它特征，计算偏移量：

1
2
3
4
5
g = data.groupby(f)
name = '_'.join(f)
for n in [1, 5, 10, 20]:
    data[f'magic_{name}_aft_gap{n}'] = g['article_id'].shift(-n) - g['article_id'].shift(0)
    data[f'magic_{name}_bef_gap{n}'] = g['article_id'].shift(0) - g['article_id'].shift(n)

还有许多特征衍生操作，这里不再一一列举，完整代码见文末。

常规特征

这一部分提取了一些字段的当日，当周以及全局的统计特征，以及历史销量滑窗的统计特征等：
Alt text

历史销量滑窗的统计特征:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
def gen_his_orders_stats(data, stride):
    orders_mean_feats = []
    orders_cols = ['author', 'level1', 'level2', 'level3', 'level4',
                   'brand', 'mall', 'url', 'baike_id_1h', 'baike_id_2h']
    for date in tqdm(range(4, 124, 7)):
        his_data = data[(data['date']<=date)&(data['date']>date-stride)].copy()
        df = data[(data['date']>date)&(data['date']<=date+7)][['article_id', 'date']+orders_cols].copy()
        for col in orders_cols:
            gp = his_data.groupby(col)['orders_3h_15h'].agg([[f'{col}_orders_{stride}d_mean', 'mean'],
                                                             [f'{col}_orders_{stride}d_std', 'std'],
                                                             [f'{col}_orders_{stride}d_min', 'min'],
                                                             [f'{col}_orders_{stride}d_max', 'max'],
                                                             [f'{col}_orders_{stride}d_med', 'median']]).reset_index()
            df = df.merge(gp, on=col, how='left')
        orders_mean_feats.append(df)
    orders_mean_feats = pd.concat(orders_mean_feats, axis=0, ignore_index=True)
    orders_mean_feats.drop(orders_cols, axis=1, inplace=True)#后面与其它表拼接，这里去除与其它表重复的字段，避免重复。
    return orders_mean_feats

调用时：

1
2
3
4
strides = [7, 150] # 150是指历史全部
for d in strides:
    feats = gen_his_orders_stats(data, d)
    data = data.merge(feats, on=['article_id', 'date'], how='left')

数据建模

训练了树模型以及深度学习模型，并将结果进行加权融合
Alt text

总结

对于本题来说，模型以及调参部分，大家都大差不差，分水岭在于特征工程，因此很有必要加强自己的特征工程能力(pandas+业务理解)，这样才有可能在下个赛场冲进前排。

最后奉上完整代码：

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
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
# -*- coding：utf-8 -*-
# Author: Ethan
# Time: 2021/8/31 14:20

import pandas as pd
import numpy as np
import warnings
import lightgbm as lgb
import os
from sklearn.model_selection import KFold, StratifiedKFold
from sklearn.preprocessing import LabelEncoder
import gc
from sklearn.metrics import mean_squared_error
from tqdm import tqdm
import seaborn as sns
import matplotlib.pyplot as plt
from catboost import CatBoostRegressor, Pool
import xgboost as xgb
warnings.filterwarnings('ignore')

def load_dataset(path):
    train_df = pd.read_csv(path+'train.csv')
    test_df = pd.read_csv(path+'test.csv')
    return train_df, test_df

def gen_his_orders_stats(data, stride):
    orders_mean_feats = []
    orders_cols = ['author', 'level1', 'level2', 'level3', 'level4',
                   'brand', 'mall', 'url', 'baike_id_1h', 'baike_id_2h']
    for date in tqdm(range(4, 124, 7)):
        his_data = data[(data['date']<=date)&(data['date']>date-stride)].copy()
        df = data[(data['date']>date)&(data['date']<=date+7)][['article_id', 'date']+orders_cols].copy()
        for col in orders_cols:
            gp = his_data.groupby(col)['orders_3h_15h'].agg([[f'{col}_orders_{stride}d_mean', 'mean'],
                                                             [f'{col}_orders_{stride}d_std', 'std'],
                                                             [f'{col}_orders_{stride}d_min', 'min'],
                                                             [f'{col}_orders_{stride}d_max', 'max'],
                                                             [f'{col}_orders_{stride}d_med', 'median']]).reset_index()
            df = df.merge(gp, on=col, how='left')
        orders_mean_feats.append(df)
    orders_mean_feats = pd.concat(orders_mean_feats, axis=0, ignore_index=True)
    orders_mean_feats.drop(orders_cols, axis=1, inplace=True)#后面与其它表拼接，这里去除与其它表重复的字段，避免重复字段。
    return orders_mean_feats

def gen_magic_features(data):
    data['magic_rank'] = data.groupby(['date'])['article_id'].rank()

    for f in tqdm([['author'],
                   ['brand'],
                   ['mall'],
                   ['price'],
                   ['level1'],
                   ['level2'],
                   ['price'],
                   ['author', 'brand']]):
        g = data.groupby(f)
        name = '_'.join(f)
        for n in [1, 5, 10, 20]:
            data[f'magic_{name}_aft_gap{n}'] = g['article_id'].shift(-n) - g['article_id'].shift(0)
            data[f'magic_{name}_bef_gap{n}'] = g['article_id'].shift(0) - g['article_id'].shift(n)
        for stat in ['max', 'min', 'mean']:
            data[f'magic_{name}_id_{stat}'] = g['article_id'].transform(stat)

    data = data.sort_values(by=['date', 'magic_rank']).reset_index(drop=True)
    for key in ['author']:
        for col in ['orders_1h', 'orders_2h', 'magic_rank']:
            data[f'{key}_{col}_bef'] = data.groupby([key, 'date'])[col].shift(1)
            data[f'{key}_{col}_aft'] = data.groupby([key, 'date'])[col].shift(-1)
            data[f'{key}_{col}_bef_diff'] = data[col] - data[f'{key}_{col}_bef']
            data[f'{key}_{col}_aft_diff'] = data[f'{key}_{col}_aft'] - data[col]

    for key in ['author', 'level1', 'level2', 'level3', 'level4',
                'brand', 'mall', 'url', 'baike_id_2h']:
        data[f'{key}_rank_mean'] = data.groupby([key, 'date'])['magic_rank'].transform('mean')
        data[f'{key}_rank_std'] = data.groupby([key, 'date'])['magic_rank'].transform('std')
        data[f'{key}_rank_min'] = data.groupby([key, 'date'])['magic_rank'].transform('min')
        data[f'{key}_rank_max'] = data.groupby([key, 'date'])['magic_rank'].transform('max')

    data['baike_id_group'] = data.groupby(['date', 'baike_id_2h'])['article_id'].rank(method='first')
    data['baike_id_group'] = data['magic_rank'] - data['baike_id_group']
    data['baike_id_group'] = data.groupby(['date', 'baike_id_2h'])['baike_id_group'].rank(ascending=False, method='dense')
    return data

def feature_extractor(train_df, test_df):
    data = pd.concat([train_df, test_df], axis=0, ignore_index=True)
    data['week'] = (data['date'] - 5) // 7 # 这里把valid和test放入到各自相同的周内
    data['price_diff_ratio'] = data['price_diff'] / (data['price'] + data['price_diff'])

    data['zhi_1h_cnt'] = data['zhi_1h'] + data['buzhi_1h']
    data['zhi_2h_cnt'] = data['zhi_2h'] + data['buzhi_2h']
    data['zhi_1h_ratio'] = data['zhi_1h'] / data['zhi_1h_cnt']
    data.loc[data['zhi_1h_cnt'] == 0, 'zhi_1h_ratio'] = 0
    data['buzhi_1h_ratio'] = data['buzhi_1h'] / data['zhi_1h_cnt']
    data.loc[data['zhi_1h_cnt'] == 0, 'buzhi_1h_ratio'] = 0
    data['zhi_2h_ratio'] = data['zhi_2h'] / data['zhi_2h_cnt']
    data.loc[data['zhi_2h_cnt'] == 0, 'zhi_2h_ratio'] = 0
    data['buzhi_2h_ratio'] = data['buzhi_2h'] / data['zhi_2h_cnt']
    data.loc[data['zhi_2h_cnt'] == 0, 'buzhi_2h_ratio'] = 0

    ### 1. magic特征
    data = gen_magic_features(data)

    ### 2. 历史销量滑窗统计特征
    strides = [7, 150] # 150是指历史全部
    for d in strides:
        feats = gen_his_orders_stats(data, d)
        data = data.merge(feats, on=['article_id', 'date'], how='left')
        del feats
        gc.collect()

    ### 3. 当日、全局count统计特征
    stats_col = ['author', 'level1', 'level2', 'level3', 'level4',
                  'brand', 'mall', 'url', 'baike_id_2h']
    for col in tqdm(stats_col):
        data[f'{col}_day_cnt'] = data.groupby([col, 'date'])['article_id'].transform('count')
        data[f'{col}_global_cnt'] = data.groupby([col])['article_id'].transform('count')
        data[f'{col}_day_cnt_ratio'] = data[f'{col}_day_cnt'] / data[f'{col}_global_cnt']

    ### 4. 当日、当周、全局orders_1h、orders_2h统计特征
    stats_col = ['author', 'level1', 'level2', 'level3', 'level4',
                 'brand', 'mall', 'url', 'baike_id_2h']
    for label in ['orders_1h', 'orders_2h']:
        for col in tqdm(stats_col):
            data[f'{col}_day_{label}_sum'] = data.groupby([col, 'date'])[label].transform('sum')
            data[f'{col}_global_{label}_sum'] = data.groupby([col])[label].transform('sum')
            data[f'{col}_day_{label}_ratio'] = data[f'{col}_day_{label}_sum'] / data[f'{col}_global_{label}_sum']
            data[f'{col}_day_{label}_mean'] = data.groupby([col, 'date'])[label].transform('mean')
            data[f'{col}_vs_day_{label}'] = data[label] / data[f'{col}_day_{label}_mean']
            data[f'{col}_week_{label}_sum'] = data.groupby([col, 'week'])[label].transform('sum')
            data[f'{col}_day_in_week_{label}_ratio'] = data[f'{col}_day_{label}_sum'] / data[f'{col}_week_{label}_sum']
            data[f'{col}_week_{label}_ratio'] = data[f'{col}_week_{label}_sum'] / data[f'{col}_global_{label}_sum']

    ### 5. 当日、当周、全局收藏、点赞、评论统计特征
    stats_col = ['author', 'url', 'baike_id_1h', 'baike_id_2h']
    for label in ['comments_1h', 'zhi_1h', 'buzhi_1h', 'favorite_1h',
                  'comments_2h', 'zhi_2h', 'buzhi_2h', 'favorite_2h']:
        for col in tqdm(stats_col):
            data[f'{col}_day_{label}_sum'] = data.groupby([col, 'date'])[label].transform('sum')
            data[f'{col}_week_{label}_sum'] = data.groupby([col, 'week'])[label].transform('sum')
            data[f'{col}_global_{label}_sum'] = data.groupby([col])[label].transform('sum')
            data[f'{col}_day_in_week_{label}_ratio'] = data[f'{col}_day_{label}_sum'] / data[f'{col}_week_{label}_sum']
            data[f'{col}_week_{label}_ratio'] = data[f'{col}_week_{label}_sum'] / data[f'{col}_global_{label}_sum']
            data[f'{col}_day_{label}_ratio'] = data[f'{col}_day_{label}_sum'] / data[f'{col}_global_{label}_sum']

    ### 6. 商品价格、降价幅度统计特征
    stats_col = ['baike_id_2h', 'url']
    for key in tqdm(stats_col):
        for col in ['price', 'price_diff', 'price_diff_ratio']:
            data[f'{key}_{col}_mean'] = data.groupby([key])[col].transform('mean')
            data[f'{key}_{col}_std'] = data.groupby([key])[col].transform('std')
            data[f'{key}_{col}_vs_mean'] = data[col] / data[f'{key}_{col}_mean']

    ### 7. 当日、当周、全局nunique统计特征
    stats_col = [['author', 'url'], ['author', 'baike_id_2h'], ['baike_id_2h', 'url']]
    for col in tqdm(stats_col):
        data[f'{col[0]}_{col[1]}_nunique_day'] = data.groupby([col[0], 'date'])[col[1]].transform('nunique')
        data[f'{col[0]}_{col[1]}_nunique_week'] = data.groupby([col[0], 'week'])[col[1]].transform('nunique')
        data[f'{col[0]}_{col[1]}_nunique_global'] = data.groupby([col[0]])[col[1]].transform('nunique')

        data[f'{col[0]}_{col[1]}_nunique_day'] = data[f'{col[0]}_{col[1]}_nunique_day'] / data[f'{col[0]}_{col[1]}_nunique_global']
        data[f'{col[0]}_{col[1]}_nunique_week'] = data[f'{col[0]}_{col[1]}_nunique_week'] / data[f'{col[0]}_{col[1]}_nunique_global']

        data[f'{col[1]}_{col[0]}_nunique_day'] = data.groupby([col[1], 'date'])[col[0]].transform('nunique')
        data[f'{col[1]}_{col[0]}_nunique_week'] = data.groupby([col[1], 'week'])[col[0]].transform('nunique')
        data[f'{col[1]}_{col[0]}_nunique_global'] = data.groupby([col[1]])[col[0]].transform('nunique')

        data[f'{col[1]}_{col[0]}_nunique_day'] = data[f'{col[1]}_{col[0]}_nunique_day'] / data[f'{col[1]}_{col[0]}_nunique_global']
        data[f'{col[1]}_{col[0]}_nunique_week'] = data[f'{col[1]}_{col[0]}_nunique_week'] / data[f'{col[1]}_{col[0]}_nunique_global']

    return data

def split_dataset(data, valid_start_date, gap):
    train = data[data['date']<valid_start_date].copy()
    valid = data[(data['date']>=valid_start_date)&(data['date']<valid_start_date+gap)].copy()
    test = data[data['date']>=valid_start_date+gap].copy()
    print(f'训练集样本数: {train.shape[0]}, 验证集样本数: {valid.shape[0]}, 测试集样本数: {test.shape[0]}')
    return train, valid, test

def lgb_reg_model(train, valid, test, target):
    params = {
        'learning_rate': 0.05,
        'boosting_type': 'gbdt',
        'objective': 'regression',
        'metric': 'l2',
        'feature_fraction': 0.8,
        'bagging_fraction': 0.8,
        'min_data_in_leaf': 200,
        'verbose': -1,
        'nthread': 16,
    }
    feats = [f for f in train.columns if f not in ['article_id', 'date', 'orders_3h_15h', 'week']]

    print('线下训练集样本数：%d' % (train.shape[0]))
    print('线下验证集样本数：%d' % (valid.shape[0]))
    print('特征个数：%d' % (len(feats)))
    print('开始线下验证......')

    dtrain = lgb.Dataset(train[feats],
                         label=train[target])
    dvalid = lgb.Dataset(valid[feats],
                         label=valid[target])
    lgb_model = lgb.train(params,
                          dtrain,
                          num_boost_round=10000,
                          valid_sets=[dvalid],
                          early_stopping_rounds=100,
                          verbose_eval=100)
    valid_preds = lgb_model.predict(valid[feats].values, num_iteration=lgb_model.best_iteration)
    print('--------------------------特征重要性----------------------------')
    feature_importance = pd.DataFrame()
    feature_importance['features'] = feats
    feature_importance['importance'] = lgb_model.feature_importance(importance_type='gain')
    print(feature_importance.sort_values(by='importance', ascending=False).head(15))
    del dtrain, dvalid
    gc.collect()

    preds = 0
    print('开始线上预测...')
    best_iters = int(lgb_model.best_iteration*1.2)
    all_train = pd.concat([train, valid], axis=0, ignore_index=True)
    dtrain = lgb.Dataset(all_train[feats],
                         label=all_train[target])
    del all_train
    gc.collect()
    lgb_model = lgb.train(params,
                          dtrain,
                          num_boost_round=best_iters,
                          verbose_eval=100,
                          )
    preds = lgb_model.predict(test[feats])
    return valid_preds, preds

def cbt_reg_model(train, valid, test, target):
    catboost_params = {
        'iterations': 10000,
        'learning_rate': 0.05,
        'eval_metric': 'RMSE',
        'task_type': 'GPU',
        'early_stopping_rounds': 200,
        'use_best_model': True,
        'verbose': 100,
        'depth': 10,
        'bootstrap_type': 'Bernoulli',
        'boosting_type': 'Plain',
        'subsample': 0.8
    }
    feats = [f for f in train.columns if f not in ['article_id', 'date', 'orders_3h_15h', 'week']]

    print('线下训练集样本数：%d' % (train.shape[0]))
    print('线下验证集样本数：%d' % (valid.shape[0]))
    print('特征个数：%d' % (len(feats)))
    print('开始线下验证......')
    dtrain = Pool(train[feats].values,
                  label=train[target])
    dvalid = Pool(valid[feats].values,
                  label=valid[target])
    cbt_model = CatBoostRegressor(**catboost_params)
    cbt_model.fit(dtrain, eval_set=dvalid)
    #     print(cbt_model.best_score_['validation'])
    score = cbt_model.best_score_['validation']['RMSE']
    valid_preds = cbt_model.predict(valid[feats].values)
    del dtrain, dvalid
    gc.collect()

    print('开始线上预测......')
    best_iters = int(cbt_model.best_iteration_ * 1.2)
    catboost_params = {
        'iterations': best_iters,
        'learning_rate': 0.05,
        'eval_metric': 'RMSE',
        'task_type': 'GPU',
        'early_stopping_rounds': 200,
        #         'use_best_model': True,
        'verbose': 100,
        'depth': 10,
        'bootstrap_type': 'Bernoulli',
        'boosting_type': 'Plain',
        'subsample': 0.8
    }
    all_train = pd.concat([train, valid], axis=0, ignore_index=True)

    dtrain = Pool(all_train[feats],
                  label=all_train[target])
    del all_train
    gc.collect()
    cbt_model = CatBoostRegressor(**catboost_params)
    cbt_model.fit(dtrain)
    del dtrain
    gc.collect()
    preds = cbt_model.predict(test[feats].values)
    return valid_preds, preds

def xgb_reg_model(train, valid, test, target):
    feats = [f for f in train.columns if f not in ['article_id', 'date', 'orders_3h_15h',
                                                   'week', 'baike_id_2h_price_diff_vs_mean', 'baike_id_2h_price_diff_ratio_vs_mean',
                                                   'url_price_diff_vs_mean', 'url_price_diff_ratio_vs_mean']]

    print('线下训练集样本数：%d' % (train.shape[0]))
    print('线下验证集样本数：%d' % (valid.shape[0]))
    print('特征个数：%d' % (len(feats)))
    print('开始线下验证......')
    clf = xgb.XGBRegressor(
        n_estimators=10000,
        max_depth=12,
        learning_rate=0.02,
        subsample=0.8,
        colsample_bytree=0.4,
        missing=-1,
        eval_metric='rmse',
        # USE CPU
        # nthread=4,
        # tree_method='hist'
        # USE GPU
        tree_method='gpu_hist'
    )
    xgb_model = clf.fit(train[feats], train[target],
                        eval_set=[(valid[feats], valid[target])],
                        verbose=100,
                        early_stopping_rounds=200)
    valid_preds = xgb_model.predict(valid[feats].values)
    #     del dtrain, dvalid
    gc.collect()

    print('开始线上预测......')
    best_iters = int(xgb_model.best_iteration * 1.2)
    clf = xgb.XGBRegressor(
        n_estimators=best_iters,
        max_depth=12,
        learning_rate=0.02,
        subsample=0.8,
        colsample_bytree=0.4,
        missing=-1,
        eval_metric='rmse',
        # USE CPU
        # nthread=4,
        # tree_method='hist'
        # USE GPU
        tree_method='gpu_hist'
    )
    all_train = pd.concat([train, valid], axis=0, ignore_index=True)

    xgb_model = clf.fit(all_train[feats],
                        all_train[target],
                        verbose=100)
    del all_train
    gc.collect()
    preds = xgb_model.predict(test[feats].values)
    return valid_preds, preds

def gen_submission(df, preds, sub_path):
    sub = df[['article_id']].copy()
    sub['orders_3h_15h'] = preds
    sub['orders_3h_15h'] = sub['orders_3h_15h'].apply(lambda x: 0 if x < 0 else x)
    sub.to_csv(sub_path+'sub.csv', index=False)

def main(data_path, sub_path):
    train_df, test_df = load_dataset(data_path) # 读取数据
    data = feature_extractor(train_df, test_df) # 生成特征
    train, valid, test = split_dataset(data=data, valid_start_date=110, gap=7) # 划分训练\验证集
    del train_df, test_df
    gc.collect()
    lgb_oof, lgb_preds = lgb_reg_model(train, valid, test, 'orders_3h_15h') # lightgbm
    cbt_oof, cbt_preds = cbt_reg_model(train, valid, test, 'orders_3h_15h')  # catboost
    xgb_oof, xgb_preds = xgb_reg_model(train, valid, test, 'orders_3h_15h')  # xgboost
    preds = 0.3*lgb_preds + 0.5*cbt_preds + 0.2*xgb_preds # 加权融合
    gen_submission(test, preds, sub_path) # 生成提交文件

if __name__ == '__main__':
    data_path = '../data/'
    sub_path = '../submission/'
    main(data_path, sub_path)

参考：