-
Notifications
You must be signed in to change notification settings - Fork 84
/
tool.py
621 lines (541 loc) · 22.2 KB
/
tool.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
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
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
import gc
import os
import sys
import time
import pickle
import datetime
import numpy as np
import pandas as pd
import xgboost as xgb
from tqdm import tqdm
import multiprocessing
import lightgbm as lgb
from scipy import stats
from functools import partial
from dateutil.parser import parse
from lightgbm import LGBMClassifier
from collections import defaultdict
from sklearn.metrics import f1_score
from datetime import date, timedelta
from contextlib import contextmanager
from sklearn.metrics import recall_score
from sklearn.cross_validation import KFold
from sklearn.metrics import precision_score
from joblib import dump, load, Parallel, delayed
from sklearn.linear_model import LinearRegression
from sklearn.model_selection import StratifiedKFold, GroupKFold
from sklearn.decomposition import NMF, PCA, TruncatedSVD
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.decomposition import LatentDirichletAllocation as LDA
# 求解rmse的均值和标准差
def get_ave_std(c1,c2,f1,f2):
'''
:param c1: 提交的常数1
:param c2: 提交的常数2
:param f1: 得分1
:param f2: 得分2
:return: 均值和标准差
'''
f1 = f1**2; f2 = f2**2;
a = 2; b = 2*(c1+c2); c = c1**2+c2**2-(f1-f2);
ave = (f1 - f2 + c2 ** 2 - c1 ** 2) / 2 / (c2 - c1)
std = (f1 - (c1 - ave) ** 2) ** 0.5
return ave,std
# 求解rmse的均值
def get_sub_ave_std(c1,c2,f1,f2,n1,n2):
'''
:param c1: 提交1的常数
:param c2: 提交2有差异的部分的常数
:param f1: 提交1的分数
:param f2: 提交2的分数
:param n1: 提交总个数
:param n2: 提交2有差异部分的个数
:return: 提交2有差异部分的均值
'''
result = ((c1+c2)-((f1**2-f2**2)*n1/n2/(c1-c2)))/2
return result
# 抽样函数
def make_sample(n,n_sub=2,seed=None):
import random
if seed is not None:
random.seed(seed)
if type(n) is int:
l = list(range(n))
s = int(n / n_sub)
else:
l = list(n)
s = int(len(n) / n_sub)
random.shuffle(l)
result = []
for i in range(n_sub):
if i == n_sub:
result.append(l[i*s:])
else:
result.append(l[i*s: (i+1)*s])
return result
# 统计list的value_counts
def value_counts(l):
s = set(l)
d = dict([(x,0) for x in s])
for i in l:
d[i] += 1
result = pd.Series(d)
result.sort_values(ascending=False,inplace=True)
return result
# 分类特征转化率
def analyse(data,name,label='label'):
result = data.groupby(name)[label].agg({'count':'count',
'sum':'sum'})
result['rate'] = result['sum']/result['count']
return result
# 连续特征转化率，等距分隔
def analyse2(data,name='id',label='label', factor=10):
grouping = pd.cut(data[name],factor)
rate = data.groupby(grouping)[label].agg({'sum':'sum',
'count':'count'})
rate['rate'] = rate['sum']/rate['count']
return rate
# 连续特征转化率，等数分隔
def analyse3(data,name='id',label='label', factor=10):
grouping = pd.qcut(data[name],factor)
rate = data.groupby(grouping)[label].agg({'sum':'sum',
'count':'count'})
rate['rate'] = rate['sum']/rate['count']
return rate
# 分组标准化
def grp_standard(data,key,names,drop=False):
for name in names:
new_name = name if drop else name + '_' + key + '_' + 'standardize'
mean_std = data.groupby(key, as_index=False)[name].agg({'mean': 'mean',
'std': 'std'})
data = data.merge(mean_std, on=key, how='left')
data[new_name] = ((data[name]-data['mean'])/data['std']).fillna(0).astype(np.float32)
data[new_name] = data[new_name].replace(-np.inf, 0).fillna(0)
data.drop(['mean','std'],axis=1,inplace=True)
return data
# 分组归一化
def grp_normalize(data,key,names,start=0,drop=False):
for name in names:
new_name = name if drop else name + '_' + key + '_' + 'normalize'
max_min = data.groupby(key,as_index=False)[name].agg({'max':'max',
'min':'min'})
data = data.merge(max_min, on=key, how='left')
data[new_name] = (data[name]-data['min'])/(data['max']-data['min'])
data[new_name] = data[new_name].replace(-np.inf, start).fillna(start).astype(np.float32)
data.drop(['max','min'],axis=1,inplace=True)
return data
# 分组排序
def grp_rank(data,key,names,ascending=True):
for name in names:
data.sort_values([key, name], inplace=True, ascending=ascending)
data['rank'] = range(data.shape[0])
min_rank = data.groupby(key, as_index=False)['rank'].agg({'min_rank': 'min'})
data = pd.merge(data, min_rank, on=key, how='left')
data['rank'] = data['rank'] - data['min_rank']
data[names] = data['rank']
data.drop(['rank'],axis=1,inplace=True)
return data
# 合并节约内存
def concat(L):
result = None
for l in L:
if result is None:
result = l
else:
result[l.columns.tolist()] = l
return result
# 分组排序函数
def group_rank(data, key, values, ascending=True):
if type(key)==list:
data_temp = data[key + [values]].copy()
data_temp.sort_values(key + [values], inplace=True, ascending=ascending)
data_temp['rank'] = range(data_temp.shape[0])
min_rank = data_temp.groupby(key,as_index=False)['rank'].agg({'min_rank':'min'})
index = data_temp.index
data_temp = data_temp.merge(min_rank,on=key,how='left')
data_temp.index = index
else:
data_temp = data[[key,values]].copy()
data_temp.sort_values(key + [values], inplace=True, ascending=ascending)
data_temp['rank'] = range(data_temp.shape[0])
data_temp['min_rank'] = data_temp[key].map(data_temp.groupby(key)['rank'].min())
data_temp['rank'] = data_temp['rank'] - data_temp['min_rank']
return data_temp['rank']
def nunique(x):
return len(set(x))
# 前后时间差的函数：
def group_diff_time(data,key,value,n):
data_temp = data[key+[value]].copy()
shift_value = data_temp.groupby(key)[value].shift(n)
data_temp['shift_value'] = data_temp[value] - shift_value
return data_temp['shift_value']
# smape
def smape(y_true,y_pred):
y_true = np.array(y_true)
y_pred = np.array(y_pred)
y_diff = np.abs(y_true-y_pred)
y_sum = y_true+y_pred
return np.mean(y_diff/y_sum)*2
# groupby 直接拼接
def groupby(data,stat,key,value,func):
key = key if type(key)==list else [key]
data_temp = data[key].copy()
feat = stat.groupby(key,as_index=False)[value].agg({'feat':func})
data_temp = data_temp.merge(feat,on=key,how='left')
return data_temp['feat']
# 计算关系紧密程度指数
def tfidf2(df,key1,key2):
key = key1 + key2
tfidf2 = '_'.join(key) + '_tfidf2'
df1 = df.groupby(key,as_index=False)[key[0]].agg({'key_count': 'size'})
df2 = df1.groupby(key1,as_index=False)['count'].agg({'key1_count': 'sum'})
df3 = df1.groupby(key2, as_index=False)['count'].agg({'key2_count': 'sum'})
df1 = df1.merge(df2,on=key1,how='left').merge(df3,on=key2,how='left')
df1[tfidf2] = df1['key_count'] / df['key2_count'] / df['key1_count']
# 相差的日期数
def diff_of_days(day1, day2):
days = (parse(day1[:10]) - parse(day2[:10])).days
return days
# 相差的分钟数
def diff_of_minutes(time1,time2):
minutes = (parse(time1) - parse(time2)).total_seconds()//60
return abs(minutes)
# 相差的小时数
def diff_of_hours(time1,time2):
hours = (parse(time1) - parse(time2)).total_seconds()//3600
return abs(hours)
# 日期的加减
def date_add_days(start_date, days):
end_date = parse(start_date[:10]) + timedelta(days=days)
end_date = end_date.strftime('%Y-%m-%d')
return end_date
# 日期的加减
def date_add_hours(start_date, hours):
end_date = parse(start_date) + timedelta(hours=hours)
end_date = end_date.strftime('%Y-%m-%d %H:%M:%S')
return end_date
# 获取某个类型里面第n次的值
def get_last_values(data, stat, key, sort_value, value, shift, sort=None):
key = key if type(key)==list else [key]
if sort == 'ascending':
stat_temp = stat.sort_values(sort_value, ascending=True)
elif sort == 'descending':
stat_temp = stat.sort_values(sort_value, ascending=False)
else:
stat_temp = stat.copy()
stat_temp['value'] = stat_temp.groupby(key)[value].shift(shift)
stat_temp.drop_duplicates(key,keep='last',inplace=True)
data_temp = data[key].copy()
data_temp = data_temp.merge(stat_temp,on=key,how='left')
return data_temp['value']
# 获取某个类型里面第n次的值
def get_first_values(data, stat, key, sort_value, value, shift, sort=None):
key = key if type(key)==list else [key]
if sort == 'ascending':
stat_temp = stat.sort_values(sort_value, ascending=True)
elif sort == 'descending':
stat_temp = stat.sort_values(sort_value, ascending=False)
else:
stat_temp = stat.copy()
stat_temp['value'] = stat_temp.groupby(key)[value].shift(-shift)
stat_temp.drop_duplicates(key,keep='first',inplace=True)
data_temp = data[key].copy()
data_temp = data_temp.merge(stat_temp,on=key,how='left')
return data_temp['value']
# 压缩数据
def compress(data):
size = sys.getsizeof(data)/2**20
def intcp(series):
ma = max(series)
mi = min(series)
if (ma<128) & (mi>=-128):
return 'int8'
elif (ma<32768) & (mi>=-32768):
return 'int16'
elif (ma<2147483648) & (mi>=-2147483648):
return 'int32'
else:
return None
def floatcp(series):
ma = max(series)
mi = min(series)
if (ma<32770) & (mi>-32770):
return 'float16'
elif (ma<2147483600) & (mi>-2147483600):
return 'float32'
else:
return None
for c in data.columns:
ctype = None
dtypes = data[c].dtypes
if dtypes == np.int64:
ctype = intcp(data[c])
if dtypes == np.int32:
ctype = intcp(data[c])
if dtypes == np.int16:
ctype = intcp(data[c])
if dtypes == np.float64:
ctype = floatcp(data[c])
if dtypes == np.float32:
ctype = floatcp(data[c])
if ctype is None:
continue
try:
data[c] = data[c].astype(ctype)
#print('{} convet to {}, done! {}'.format(dtypes,ctype,c))
except:
print('error')
#print('特征{}的类型为：{}，转化出线问题！！！'.format(c,dtypes))
#print('原始数据大小为： {}M'.format(round(size, 2)))
#print('新数据大小为： {}M'.format(round(sys.getsizeof(data) / 2 ** 20,2)))
return data
def trend(y):
try:
x = np.arange(0, len(y)).reshape(-1, 1)
lr = LinearRegression()
lr.fit(x, y)
trend = lr.coef_[0]
except:
trend = np.nan
return trend
@contextmanager
def timer(title):
t0 = time.time()
yield
#print("{} - done in {:.0f}s".format(title, time.time() - t0))
def jiangwei(stat,data, id, feature):
#print('lda ...')
mapping = {}
for sample in stat[[id, feature]].values:
mapping.setdefault(sample[0], []).append(str(sample[1]))
ids = list(mapping.keys())
sentences = [' '.join(mapping[cate_]) for cate_ in ids]
stat_sentences_matrix = CountVectorizer(token_pattern='(?u)\\b\\w+\\b', min_df=2).fit_transform(sentences)
mapping = {}
for sample in data[[id, feature]].values:
mapping.setdefault(sample[0], []).append(str(sample[1]))
ids = list(mapping.keys())
sentences = [' '.join(mapping[cate_]) for cate_ in ids]
data_sentences_matrix = CountVectorizer(token_pattern='(?u)\\b\\w+\\b', min_df=2).fit_transform(sentences)
lda = LDA(n_components=5,
learning_method='online',
batch_size=1000,
n_jobs=40,
random_state=520)
lda.fit(stat_sentences_matrix)
lda_matrix = lda.transform(data_sentences_matrix)
lda_matrix = pd.DataFrame(lda_matrix,columns=['lda_{}_{}'.format(feature, i) for i in range(5)]).astype('float16')
nmf = NMF(n_components=5,
random_state=520,
beta_loss='kullback-leibler',
solver='mu',
max_iter=1000,
alpha=.1,
l1_ratio=.5)
nmf.fit(stat_sentences_matrix)
nmf_matrix = nmf.transform(stat_sentences_matrix)
nmf_matrix = pd.DataFrame(nmf_matrix,columns=['nmf_{}_{}'.format(feature, i) for i in range(5)]).astype('float16')
pca = TruncatedSVD(5)
pca.fit(stat_sentences_matrix)
pca_matrix = pca.transform(stat_sentences_matrix)
pca_matrix = pd.DataFrame(pca_matrix,
columns=["%s_%s_svd_action" % ('user_sku', i) for i in range(5)]).astype('float32')
matrix = concat([lda_matrix,nmf_matrix,pca_matrix])
matrix[id] = ids
return matrix
# 获取阈值
def get_threshold(preds,silent=False):
preds_temp = sorted(preds,reverse=True)
n = sum(preds) # 实际正例个数
m = 0 # 提交的正例个数
e = 0 # 正确个数的期望值
f1 = 0 # f1的期望得分
for threshold in preds_temp:
e += threshold
m += 1
f1_temp = e/(m+n)
if f1>f1_temp:
break
else:
f1 = f1_temp
if not silent:
print('error')
#print('阈值为：{}'.format(threshold))
#print('提交正例个数为：{}'.format(m-1))
#print('期望得分为：{}'.format(f1*2))
return [(1 if (pred>threshold) else 0) for pred in preds]
# 多分类F1值
def multi_f1(true,pred,silent=False):
true_dummy = pd.get_dummies(pd.Series(true))
pred_dummy = pd.get_dummies(pd.Series(pred))
scores = []
for c in true_dummy.columns:
score = f1_score(true_dummy[c],pred_dummy[c])
if not silent:
pass
#print('{} : {}'.format(c,score))
scores.append(score)
return np.mean(scores)
# 多分类f1期望得分
def exp_multi_f1(pred,int_preds,silent=False):
int_preds_dummy = pd.get_dummies(pd.Series(int_preds))
pred = pd.DataFrame(pred,columns=int_preds_dummy.columns)
scores = []
for c in pred.columns:
n = pred[c].sum()
m = int_preds_dummy[c].sum()
r = pred[int_preds_dummy[c]==1][c].sum()
f1 = 2*r / (m+n)
if not silent:
pass
#print('{} : {}'.format(c, f1))
scores.append(f1)
return np.mean(scores)
# 多分类f1最佳阈值
def get_multi_f1_threshold(preds):
def derivative(arg, p):
m, n, r = arg
s = m + n
return 2 * (p * s - r) / (s + 1) / s
def get_multi_f1_threshold_di(preds, int_preds, preds_flag):
int_preds_matrix = pd.get_dummies(int_preds).values
para_dict = {}
for i in range(preds.shape[1]):
m = preds[:, i].sum()
n = int_preds_matrix[:, i].sum()
r = preds[int_preds_matrix[:, i] == 1, i].sum()
para_dict[i] = (m, n, r)
for i in range(preds.shape[0]):
if preds_flag[i]:
continue
else:
int_preds[i] = np.argmax([derivative(para_dict[j],preds[i,j]) for j in range(preds.shape[1])])
return int_preds
int_preds = pd.Series(preds.argmax(axis=1))
preds_flag = list(preds.max(axis=1)>0.5)
for i in range(3):
int_preds = get_multi_f1_threshold_di(preds,int_preds,preds_flag)
#print('期望的分： {}'.format(exp_multi_f1(preds,int_preds)))
# 贝叶斯平滑
def bayes_encode(C, I):
def compute_moment(tries, success):
ctr_list = []
var = 0.0
for i in range(len(tries)):
ctr_list.append(float(success[i]) / tries[i])
mean = sum(ctr_list) / len(ctr_list)
for ctr in ctr_list:
var += pow(ctr - mean, 2)
return mean, var / (len(ctr_list) - 1)
def update_from_data_by_moment(tries, success):
mean, var = compute_moment(tries, success)
alpha = (mean + 0.000001) * ((mean + 0.000001) * (1.000001 - mean) / (var + 0.000001) - 1)
beta = (1.000001 - mean) * ((mean + 0.000001) * (1.000001 - mean) / (var + 0.000001) - 1)
return alpha, beta
I_temp = list(I)
C_temp = list(C)
alpha, beta = update_from_data_by_moment(I_temp, C_temp)
rate = ((alpha + C) / (alpha + beta + I)).astype('float32')
return rate
# 交叉验证统计转化率
def mul_cv_convert(data,category,label,cv=5):
rate = np.zeros((len(data),data[label].nunique()))
kf = KFold(len(data), n_folds=cv, shuffle=True, random_state=66)
for i, (train_index, test_index) in enumerate(kf):
stat1 = data.iloc[train_index]
stat2 = data.iloc[test_index]
temp1 = stat1.groupby([category, label], as_index=False).size().unstack().fillna(0)
temp2 = stat1[~stat1[label].isnull()].groupby([category], as_index=False).size()
temp3 = (temp1.T / temp2).T
columns = [category + '_' + str(c) + '_conversion' for c in temp3.columns]
temp3 = temp3.reset_index()
temp4 = stat2[[category]].merge(temp3, on=category, how='left')
rate[test_index,:] = temp4.drop(category,axis=1).values
rate = pd.DataFrame(rate,columns=columns)
data = concat([data,rate],axis=1)
return data
# count encoding
def count_encoding(li):
temp = pd.Series(li)
result = temp.map(temp.value_counts())
return result
# 众位数
def mode(li):
if stats.mode(li)[1][0]==1:
return np.nan
return stats.mode(li)[0][0]
def xgb_cv(params, train_feat, test_feat, predictors, label='label',groups=None,cv=5,stratified=True):
#print('开始CV 5折训练...')
t0 = time.time()
train_preds = np.zeros((len(train_feat), train_feat[label].nunique()))
test_preds = np.zeros((len(test_feat), train_feat[label].nunique()))
xgb_test = xgb.DMatrix(test_feat[predictors])
models = []
group_kfold = GroupKFold(n_splits=cv).get_n_splits(train_preds, train_preds[label], groups, random_state=66)
for i, train_index, test_index in group_kfold.split(train_preds, train_preds[label], groups):
xgb_train = xgb.DMatrix(train_feat[predictors].iloc[train_index], train_feat[label].iloc[train_index])
xgb_eval = xgb.DMatrix(train_feat[predictors].iloc[test_index], train_feat[label].iloc[test_index])
#print('开始第{}轮训练...'.format(i))
params = {'objective': 'multi:softprob',
'eta': 0.1,
'max_depth': 6,
'silent': 1,
'num_class': 11,
'eval_metric': "mlogloss",
'min_child_weight': 3,
'subsample': 0.7,
'colsample_bytree': 0.7,
'seed': 66
} if params is None else params
watchlist = [(xgb_train, 'train'), (xgb_eval, 'val')]
clf = xgb.train(params,
xgb_train,
num_boost_round=3000,
evals=watchlist,
verbose_eval=50,
early_stopping_rounds=50)
train_preds[test_index] += clf.predict(xgb_eval)
test_preds += clf.predict(xgb_test)
models.append(clf)
pickle.dump(models,open('xgb_{}.model'.format(datetime.datetime.now().strftime('%Y%m%d_%H%M%S')),'+wb'))
#print('用时{}秒'.format(time.time()-t0))
return train_preds,test_preds/5
def lgb_cv(params, train_feat, test_feat, predictors, label='label', cv=5,stratified=True):
#print('开始CV 5折训练...')
t0 = time.time()
train_preds = np.zeros((len(train_feat), train_feat[label].nunique()))
test_preds = np.zeros((len(test_feat), train_feat[label].nunique()))
models = []
group_kfold = GroupKFold(n_splits=cv).get_n_splits(train_preds, train_preds[label], groups, random_state=66)
for i, train_index, test_index in group_kfold.split(train_preds, train_preds[label], groups):
lgb_train = lgb.Dataset(train_feat[predictors].iloc[train_index], train_feat['label'].iloc[train_index])
lgb_eval = lgb.Dataset(train_feat[predictors].iloc[test_index], train_feat['label'].iloc[test_index])
#print('开始第{}轮训练...'.format(i))
params = {
'task': 'train',
'boosting_type': 'gbdt',
'objective': 'multiclass',
'num_class':11,
'max_depth': 8,
'num_leaves': 150,
'learning_rate': 0.05,
'subsample': 0.7,
'colsample_bytree': 0.7,
'feature_fraction': 0.9,
'bagging_fraction': 0.95,
'bagging_freq': 5,
'verbose': 0,
'seed': 66,
} if params is None else params
clf = lgb.train(params,
lgb_train,
num_boost_round=10000,
valid_sets=lgb_eval,
verbose_eval=50,
early_stopping_rounds=100)
train_preds[test_index] += clf.predict(train_feat[predictors].iloc[test_index])
test_preds += clf.predict(test_feat[predictors])
models.append(clf)
pickle.dump(models, open('xgb_{}.model'.format(datetime.datetime.now().strftime('%Y%m%d_%H%M%S')), '+wb'))
#print('用时{}秒'.format(time.time() - t0))
return train_preds, test_preds / 5