-
Notifications
You must be signed in to change notification settings - Fork 0
/
solution.py
743 lines (561 loc) · 27.7 KB
/
solution.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
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
# coding: utf-8
# # Kaggle 2017: Two Sigma Connect: Rental Listing Inquiries [Competition](https://www.kaggle.com/c/two-sigma-connect-rental-listing-inquiries)
#
# ** * Silver solution (118-th place) * **
# ** * Original submission is [here](https://github.com/vecxoz/kag17_2sigma_renthop/tree/master/original_best_submission) * **
# ** * Public LB: 0.50762 * **
# ** * Private LB: 0.50665 * **
# ** * Author: Igor Ivanov ([vecxoz](https://www.kaggle.com/vecxoz)) * **
# ** * Email: [email protected] * **
# ** * MIT License * **
# # How to Reproduce
#
# * Submission score may slightly vary depending on versions of packages, but should be around 120-th place
# * You need 8 GB RAM
# * On machine with 4 cores solution runs about 2 hours
# * To reproduce solution basically you need Ubuntu, Java and Python 3 with
# NumPy, Pandas, SciPy, Scikit-learn and XGBoost (maybe I missed something)
# * You can deploy ML environment on Ubuntu for Python 3 using this [script](https://github.com/vecxoz/vecsnip/blob/master/deploy_cloud_ml_ubuntu_python_no_gpu.sh)
# * Clone (or download) repository [https://github.com/vecxoz/kag17_2sigma_renthop](https://github.com/vecxoz/kag17_2sigma_renthop). You will have dir `kag17_2sigma_renthop`
# * Put `train.json` and `test.json` files into `kag17_2sigma_renthop/data`. You can download this files from competition [data page](https://www.kaggle.com/c/two-sigma-connect-rental-listing-inquiries/data)
# * Run `$ python3 solution.py` or just run all cells of this notebook `solution.ipynb`
# * Files `solution.py` and `solution.ipynb` contain completely identical Python code
# * After script is complete you will have submission file: `kag17_2sigma_renthop/reproduced_submission/reproduced_submission.csv`
# # Overview
# The dataset for this competition is just amazing.
# We have all kinds of features: numerical, categorical, geospatial (lat/lon), text, pictures...
# Just endless possibilities for feature engineering.
# Let's look at some training example:
# ```
# >>> train_df.iloc[12]
# bathrooms 1.000000
# bedrooms 2
# building_id 67c9b420da4a365bc26a6cd0ef4a5320
# created 2016-04-19 05:37:25
# description ***LOW FEE. Beautiful CHERRY OAK WOODEN FLOORS...
# display_address E 38th St
# features [Doorman, Elevator, Laundry in Building, No Fee]
# interest_level high
# latitude 40.748800
# listing_id 6895442
# longitude -73.977000
# manager_id 537e06890f6a86dbb70c187db5be4d55
# photos [https://photos.renthop.com/2/6895442_34d617a5...
# price 3000
# street_address 137 E 38th St
# ```
# The final submission is an ensemble (weighted average) of 3 first-level models.
# Each first-level model is meta-model by nature itself.
# First-level models are built based on the concept of 'mixed stacking':
# * fit some model on dataset
# * predict dataset
# * append predictions to dataset
# * fit some other model on dataset + predictions
#
# Algorithms used:
# * [XGBoost](https://github.com/dmlc/xgboost)
# * [StackNet](https://github.com/kaz-Anova/StackNet)
# * [Extra Trees](http://scikit-learn.org/stable/modules/generated/sklearn.ensemble.ExtraTreesClassifier.html)
# 
# # Solution
# ## Import
# In[1]:
# Basics
import os
import sys
import gc
import re
from subprocess import check_output
# Math stack
import numpy as np
np.set_printoptions(suppress = True)
import pandas as pd
# pd.set_option('display.float_format', lambda x: '%.6f' % x)
pd.options.mode.chained_assignment = None # default = 'warn'
from scipy import sparse
from scipy.optimize import minimize
# Preprocessing and scoring
from sklearn.model_selection import cross_val_score, KFold, StratifiedKFold, train_test_split
from sklearn.metrics import log_loss, make_scorer
from sklearn.preprocessing import LabelEncoder, StandardScaler
# Text (vectorizing, stamming, sentiment)
from sklearn.feature_extraction.text import CountVectorizer, TfidfVectorizer
# from textblob import TextBlob
# Models
import xgboost as xgb
from xgboost import XGBClassifier
from sklearn.ensemble import ExtraTreesClassifier
# Default value to fill NaN
fill_val = 0
# ## Load data
# In[2]:
data_dir = './data/'
train_df = pd.read_json(data_dir + 'train.json')
test_df = pd.read_json(data_dir + 'test.json')
subm_df = pd.read_csv(data_dir + 'sample_submission.csv')
# Load "magic feature"
time_df = pd.read_csv(data_dir + 'listing_image_time.csv')
# Rename columns
time_df.columns = ['listing_id', 'timestamp']
#
print(train_df.shape) # (49352, 15)
print(test_df.shape) # (74659, 14)
# ## Combine train and test to simplify feature calculation
# In[3]:
y_col = 'interest_level'
r, c = train_df.shape
test_df.loc[:, y_col] = 'na'
tt_df = pd.concat([train_df, test_df], ignore_index = True)
# Merge with "magic feature"
tt_df = pd.merge(tt_df, time_df, on = 'listing_id', how = 'left')
print(tt_df.shape) # (124011, 16)
# ## Minimal outlier correction
# In[4]:
# bathrooms
tt_df.loc[69023, 'bathrooms'] = 2 # was 112
tt_df.loc[72329, 'bathrooms'] = 2 # was 20
tt_df.loc[113071, 'bathrooms'] = 2 # was 20
tt_df.loc[1990, 'bathrooms'] = 1 # was 10
# lat/lon - just another city - e.g. LA - leave as is
# price
tt_df.loc[25538, 'price'] = 1025 # was 111111 # real number from dscription
tt_df.loc[tt_df['price'] > 100000, 'price'] = 100000 # low interest_level for all
# timestamp
tt_df.loc[35264, 'timestamp'] = 1479787252 # was 1491289977 (only one record from april) # replace with last timestamp excluding this record
# ## Feature engineering
# In[5]:
#---------------------------------------------------------------------------
# Features denoting presence of NaNs, zeros, outliers
#---------------------------------------------------------------------------
tt_df['building_id_is_zero'] = (tt_df['building_id'].apply(len) == 1).astype(np.int64)
#---------------------------------------------------------------------------
# Count/len
#---------------------------------------------------------------------------
tt_df['num_photos'] = tt_df['photos'].apply(len) # number of photos
tt_df['num_features'] = tt_df['features'].apply(len) # number of 'features'
tt_df['num_description_words'] = tt_df['description'].apply( lambda x: len(x.split(' ')) ) # number of words in description
#---------------------------------------------------------------------------
# Date/Time
#---------------------------------------------------------------------------
tt_df['created'] = pd.to_datetime(tt_df['created']) # convert the created column to datetime
# tt_df['year'] = tt_df['created'].dt.year # year is constant for this dataset
tt_df['month'] = tt_df['created'].dt.month
tt_df['day'] = tt_df['created'].dt.day
tt_df['hour'] = tt_df['created'].dt.hour
#---------------------------------------------------------------------------
# Rooms
#---------------------------------------------------------------------------
tt_df['bad_plus_bath'] = tt_df['bedrooms'] + tt_df['bathrooms']
tt_df['more_bed'] = (tt_df['bedrooms'] > tt_df['bathrooms']).astype(np.int64)
tt_df['more_bath'] = (tt_df['bedrooms'] < tt_df['bathrooms']).astype(np.int64)
tt_df['bed_bath_equal'] = (tt_df['bedrooms'] == tt_df['bathrooms']).astype(np.int64)
tt_df['bed_bath_diff'] = tt_df['bedrooms'] - tt_df['bathrooms']
tt_df['bed_bath_ration'] = tt_df['bedrooms'] / tt_df['bathrooms']
tt_df['bed_bath_ration'] = tt_df['bed_bath_ration'].replace([np.inf], np.max(tt_df.loc[tt_df['bed_bath_ration'] != np.inf, 'bed_bath_ration']) + 1)
tt_df['bed_bath_ration'].fillna(0, inplace = True)
tt_df['bath_is_int'] = (0 == tt_df['bathrooms'] % 1).astype(np.int64)
# tt_df['diff_rooms_photos'] = tt_df['num_photos'] - tt_df['bad_plus_bath']
tt_df['bed_bath_photos_ration'] = tt_df['bad_plus_bath'] / tt_df['num_photos']
tt_df['bed_bath_photos_ration'] = tt_df['bed_bath_photos_ration'].replace([np.inf], np.max(tt_df.loc[tt_df['bed_bath_photos_ration'] != np.inf, 'bed_bath_photos_ration']) + 1)
tt_df['bed_bath_photos_ration'] = tt_df['bed_bath_photos_ration'].replace([np.nan], 0)
#---------------------------------------------------------------------------
# Price
#---------------------------------------------------------------------------
tt_df['price_per_bed'] = tt_df['price'] / tt_df['bedrooms']
tt_df['price_per_bed'] = tt_df['price_per_bed'].replace([np.inf], np.max(tt_df.loc[tt_df['price_per_bed'] != np.inf, 'price_per_bed']) + 1)
tt_df['price_per_bath'] = tt_df['price'] / tt_df['bathrooms']
tt_df['price_per_bath'] = tt_df['price_per_bath'].replace([np.inf], np.max(tt_df.loc[tt_df['price_per_bath'] != np.inf, 'price_per_bath']) + 1)
tt_df['price_per_bed_plus_bath'] = tt_df['price'] / (tt_df['bedrooms'] + tt_df['bathrooms'])
tt_df['price_per_bed_plus_bath'] = tt_df['price_per_bed_plus_bath'].replace([np.inf], np.max(tt_df.loc[tt_df['price_per_bed_plus_bath'] != np.inf, 'price_per_bed_plus_bath']) + 1)
tt_df['price_per_photo'] = tt_df['price'] / tt_df['num_photos']
tt_df['price_per_photo'] = tt_df['price_per_photo'].replace([np.inf], np.max(tt_df.loc[tt_df['price_per_photo'] != np.inf, 'price_per_photo']) + 1)
#---------------------------------------------------------------------------
# Address (case may contain info)
#---------------------------------------------------------------------------
tt_df['street_address'] = tt_df['street_address'].apply(lambda x: x.lower())
tt_df['display_address'] = tt_df['display_address'].apply(lambda x: x.lower())
# tt_df['disp_addr_is_not_in_street_addr'] = tt_df[['street_address', 'display_address']].apply(lambda x: np.int(-1 == x.street_address.find(x.display_address)), axis = 1)
#---------------------------------------------------------------------------
# Lat/Lon
#---------------------------------------------------------------------------
# # latlon count (density of points)
tt_df['latlon'] = tt_df['longitude'].round(3).astype(str) + '_' + tt_df['latitude'].round(3).astype(str)
latlon_count = tt_df['latlon'].value_counts()
latlon_count = latlon_count.reset_index().rename(columns = {'index':'latlon', 'latlon':'density'})
tt_df = pd.merge(tt_df, latlon_count, on = 'latlon', how = 'left')
# Distance to New-Yourk center
center_lat = 40.785091
center_lon = -73.968285
tt_df['euclid_dist_to_center'] = np.sqrt((tt_df['latitude'] - center_lon) ** 2 + (tt_df['longitude'] - center_lat) ** 2)
# Rotation for different angles
for angle in [15,30,45,60]:
namex = 'rot' + str(angle) + '_x'
namey = 'rot' + str(angle) + '_y'
alpha = np.pi / (180 / angle)
tt_df[namex] = tt_df['latitude'] * np.cos(alpha) + tt_df['longitude'] * np.sin(alpha)
tt_df[namey] = tt_df['longitude'] * np.cos(alpha) - tt_df['latitude'] * np.sin(alpha)
#---------------------------------------------------------------------------
# Categotical
#---------------------------------------------------------------------------
# Label encoding
categorical_cols = ['display_address', 'manager_id', 'building_id', 'street_address']
for col in categorical_cols:
le = LabelEncoder()
tt_df.loc[:, col] = le.fit_transform(tt_df[col].values)
# Manager count
man_count = tt_df['manager_id'].value_counts()
man_count = man_count.reset_index().rename(columns = {'index':'manager_id', 'manager_id':'man_count'})
tt_df = pd.merge(tt_df, man_count, on = 'manager_id', how = 'left')
# Building count
build_count = tt_df['building_id'].value_counts()
build_count = build_count.reset_index().rename(columns = {'index':'building_id', 'building_id':'build_count'})
tt_df = pd.merge(tt_df, build_count, on = 'building_id', how = 'left')
# Top5 building
build_count = tt_df['building_id'].value_counts()
p = np.percentile(build_count.values, 95)
tt_df['top_5_building'] = tt_df['building_id'].apply( lambda x: np.int(x in build_count.index.values[build_count.values >= p]) )
#---------------------------------------------------------------------------
# Dscription
# Description in fact is the list of features, so probably it can add little values to 'features'
#---------------------------------------------------------------------------
tt_df['number_of_new_lines'] = tt_df['description'].apply(lambda x: x.count('<br /><br />'))
tt_df['website_redacted'] = tt_df['description'].str.contains('website_redacted').astype(np.int)
#---------------------------------------------------------------------------
# Strange
#---------------------------------------------------------------------------
tt_df['price_is_round_sousand'] = (0 == tt_df['price'] % 1000).astype(np.int64)
tt_df['price_is_round_hundred'] = (0 == tt_df['price'] % 100).astype(np.int64)
#---------------------------------------------------------------------------
# Image timestamp ('magic feature')
#---------------------------------------------------------------------------
tt_df['ts_date'] = pd.to_datetime(tt_df['timestamp'], unit = 's')
# tt_df['ts_days_passed'] = (tt_df['ts_date'].max() - tt_df['ts_date']).astype('timedelta64[D]').astype(int)
tt_df['ts_month'] = tt_df['ts_date'].dt.month
tt_df['ts_week'] = tt_df['ts_date'].dt.week
tt_df['ts_day'] = tt_df['ts_date'].dt.day
# tt_df['ts_dayofweek'] = tt_df['ts_date'].dt.dayofweek
tt_df['ts_dayofyear'] = tt_df['ts_date'].dt.dayofyear
tt_df['ts_hour'] = tt_df['ts_date'].dt.hour
tt_df['ts_tensdays'] = tt_df['ts_day'].apply(lambda x: 1 if x < 10 else 2 if x < 20 else 3)
#---------------------------------------------------------------------------
# Check NaNs
#---------------------------------------------------------------------------
print(tt_df.shape) # (124011, 60)
print('NaN: %s' % tt_df.isnull().mean().any())
# ## Split dataset to calculate features based on taraget variable (probabilities)
# In[6]:
train_df = tt_df[:r]
test_df = tt_df[r:]
# ## Features based on taraget variable (probabilities)
# ** * Perform groupping of interest_level by (manager_id and interest_level) * **
# ** * Should be very careful to avoid leakage. For example more than 5 folds will increase possibility of leakage * **
# ## Function to calculate probabilities
# In[7]:
def get_prob(df, col = None, agg_func = None):
"""
Params
------
df - Panadas dataframe
col - column of interest
agg_func - aggregation function
Return
------
Pandas dataframe ready to merge with df on manager_id
Logic
-----
We have this: We want to get this:
------------- --------------------
interest_level manager_id manager_id prob_high prob_low prob_medium
low foo bar 0.333333 NaN 0.666667
medium bar foo 0.200000 0.4 0.400000
medium foo
high bar
medium foo
medium bar
low foo
high foo
"""
aggregate_df = df.groupby(['manager_id', 'interest_level'])[[col]].aggregate(agg_func).rename(columns = {col: 'aggregate'}).reset_index()
sum_df = aggregate_df.groupby(['manager_id'])[['aggregate']].sum().rename(columns = {'aggregate': 'sum'}).reset_index()
aggregate_df = pd.merge(aggregate_df, sum_df, on = 'manager_id', how = 'left')
aggregate_df['prob'] = aggregate_df['aggregate'] / aggregate_df['sum']
piv_df = pd.pivot_table(aggregate_df, values='prob', columns=['interest_level'], index = 'manager_id').reset_index()
name = col + '_' + agg_func
piv_df.rename(columns = {'high': 'prob_high_' + name, 'low': 'prob_low_' + name, 'medium': 'prob_medium_' + name}, inplace = True)
return piv_df
# ## Calculate probabilities
# In[8]:
# Init CV
kf = KFold(n_splits = 5, shuffle = True, random_state = 0)
# Init aggregation
col = 'interest_level'
agg_func = 'count'
# Init new columns
for i in ['prob_high_', 'prob_low_', 'prob_medium_']: # alphabetically
train_df.loc[:, i + col + '_' + agg_func] = fill_val
# For train set
for train_index, test_index in kf.split(train_df):
tr_df = train_df.iloc[train_index]
te_df = train_df.iloc[test_index]
piv_df = get_prob(tr_df, col = col, agg_func = agg_func)
te_df = pd.merge(te_df, piv_df, on = 'manager_id', how = 'left')
train_df.iloc[test_index, -3:] = te_df.iloc[:, -3:].values
print('Fold done')
# For test set
piv_df = get_prob(train_df, col = col, agg_func = agg_func)
test_df = pd.merge(test_df, piv_df, on = 'manager_id', how = 'left')
# Fill NaN
train_df.fillna(fill_val, inplace = True)
test_df.fillna(fill_val, inplace = True)
print(train_df.shape) # (49352, 63)
print(test_df.shape) # (74659, 63)
# ## Encode target
# In[9]:
train_df.loc[:, 'interest_level'] = train_df['interest_level'].map({'high': 0, 'medium': 1, 'low': 2})
y_train = train_df['interest_level'].values
# ## Combine train and test to work with text features
# In[10]:
# Combine train and test
tt_df = pd.concat([train_df, test_df], ignore_index = True)
print(tt_df.shape) # (124011, 63)
# ## Create text features (sparse) from 'features' column
# In[11]:
# Text features from 'features'
tt_df['features'] = tt_df['features'].apply(lambda x: ' '.join(['_'.join(i.split(' ')) for i in x]))
vectorizer = CountVectorizer(stop_words = 'english', max_features = 200)
tt_sparse = vectorizer.fit_transform(tt_df['features'])
# ## Create text features from 'description' column (sentiment, NO improvement)
# In[12]:
# Text features from 'description'
# tt_df['sentiment_polarity'] = tt_df['description'].apply(lambda x: TextBlob(x).sentiment.polarity)
# tt_df['sentiment_subjectivity'] = tt_df['description'].apply(lambda x: TextBlob(x).sentiment.subjectivity)
# ## Create final dataset
#
# * Explicitly select columns to use
# * Combine dense and sparse data into a single sparse dataset
# * Check consistency: NaN, +INF, -INF, constant columns, duplicated columns
# In[13]:
X_cols = ['bathrooms', 'bedrooms', 'latitude', 'longitude', 'price',
'listing_id', 'num_photos', 'num_features', 'num_description_words',
'month', 'day', 'hour', 'bad_plus_bath', 'more_bed',
'more_bath', 'bed_bath_equal', 'bed_bath_diff', 'bed_bath_ration',
'price_per_bed', 'price_per_bath', 'price_per_bed_plus_bath',
'price_per_photo', 'price_is_round_sousand', 'price_is_round_hundred',
'building_id_is_zero', 'bath_is_int', 'bed_bath_photos_ration',
'density', 'euclid_dist_to_center',
'prob_high_interest_level_count', 'prob_low_interest_level_count', 'prob_medium_interest_level_count',
'display_address', 'manager_id', 'building_id', 'street_address',
'man_count', 'build_count', 'top_5_building',
'ts_month', 'ts_week', 'ts_day', 'ts_dayofyear', 'ts_hour', 'ts_tensdays',
'rot15_x', 'rot15_y', 'rot30_x', 'rot30_y',
'rot45_x', 'rot45_y', 'rot60_x', 'rot60_y',
'number_of_new_lines', 'website_redacted',
]
TT = sparse.hstack([tt_df[X_cols], tt_sparse]).tocsr()
# TT = sparse.csr_matrix(tt_df[X_cols]) # without text features
# Check for NaN, INF, -INF
print('NaN -> ', np.bool(np.mean(np.isnan(TT.toarray())))) # should be False
print('+INF -> ', np.bool(np.mean(np.isinf(TT.toarray())))) # should be False
print('-INF -> ', np.bool(np.mean(np.isneginf(TT.toarray())))) # should be False
# Check for constant fetures
print('CONST -> ', np.bool(np.mean(TT[0] == TT.mean(axis = 0)))) # should be False
# Check for duplicate entries in column (feature) list
print('DUPL -> ', len(X_cols) != len(set(X_cols))) # should be False
# Split
X_train = TT[:r]
X_test = TT[r:]
# Shape
print('SHAPE -> ', X_train.shape, X_test.shape) # (49352, 255) (74659, 255)
# ## Organize cross-validation for manual parameter tuning and feature selection
# In[14]:
# Init model
model = XGBClassifier(seed = 0, objective = 'multi:softprob',
learning_rate = 0.1, n_estimators = 100,
max_depth = 6, min_child_weight = 1,
subsample = 0.7, colsample_bytree = 0.7)
# Crate sklearn scorer
scorer = make_scorer(log_loss, needs_proba = True)
# Run CV and get mean score
print(np.mean(cross_val_score(model, X_train, y_train, cv = 3, scoring = scorer)))
# ## Function to create out-of-fold predictions
# In[15]:
def oof(model, X_train, y_train, X_test, oof_test = True):
"""
Parameters
----------
Self-explanatory
oof_test - if True, then predict test set
Return
------
S_train - OOF predictions for train set
S_test - prediction for test set (fit model on full train set)
"""
# Init CV
kf = KFold(n_splits = 3, shuffle = True, random_state = 0)
# Create empty numpy arrays for stacking features
S_train = np.zeros((X_train.shape[0], 3))
S_test = np.zeros((X_test.shape[0], 3))
# Create oof predictions for train set
for i, (train_index, test_index) in enumerate(kf.split(X_train)):
X_tr = X_train[train_index]
y_tr = y_train[train_index]
X_te = X_train[test_index]
y_te = y_train[test_index]
model = model.fit(X_tr, y_tr)
y_te_pred = model.predict_proba(X_te)
S_train[test_index, :] = y_te_pred
print( 'Fold %d: %.6f' % (i, log_loss(y_te, y_te_pred)) )
# Score over full dataset (mean)
print( 'Mean: %.6f' % log_loss(y_train, S_train) )
# Create prediction for test set (fit on full train)
if oof_test:
model = model.fit(X_train, y_train)
S_test = model.predict_proba(X_test)
return (S_train, S_test)
# # First-level model 1: XGBoost
# ## Determine number of rounds for XGBoost using native CV
# In[16]:
# Parameters
params = {'seed': 0,
'objective': 'multi:softprob',
'eval_metric': 'mlogloss',
'num_class': 3,
'eta': 0.02,
'max_depth': 6,
'min_child_weight': 1,
'subsample': 0.7,
'colsample_bytree': 0.7,
'silent': 1,
}
# Convert data to DMatrices
dtrain = xgb.DMatrix(X_train, label = y_train)
dtest = xgb.DMatrix(X_test)
# Using 3-fold CV
res = xgb.cv(
params,
dtrain,
num_boost_round = 10000,
early_stopping_rounds = 50,
nfold = 3,
seed = 0,
stratified = False,
show_stdv = True,
verbose_eval = 100
)
# Output result
n_part = res.shape[0]
n_full = np.int(res.shape[0] + (1/3) * res.shape[0])
print('\ncv mean + std -> [%.6f + %.6f]\nntrees -> [%d]\nntrees for full data (+1/3) -> [%d]' % (res.iloc[-1, 0], res.iloc[-1, 1], n_part, n_full))
# ## Run XGBoost
# In[17]:
# Init model
model = XGBClassifier(seed = 0, objective = 'multi:softprob',
learning_rate = 0.02, n_estimators = n_part,
max_depth = 6, min_child_weight = 1,
subsample = 0.7, colsample_bytree = 0.7)
# Get oof
xgb_oof_train, xgb_oof_test = oof(model, X_train, y_train, X_test, oof_test = False)
# Init model for test (as we train on full train set we need more rounds)
model = XGBClassifier(seed = 0, objective = 'multi:softprob',
learning_rate = 0.02, n_estimators = n_full,
max_depth = 6, min_child_weight = 1,
subsample = 0.7, colsample_bytree = 0.7)
# Fit model on full train
model = model.fit(X_train, y_train)
# Predict test
xgb_oof_test = model.predict_proba(X_test)
# Export to txt files
np.savetxt(data_dir + 'xgb_oof_train.csv', xgb_oof_train, delimiter = ',', fmt = '%.5f')
np.savetxt(data_dir + 'xgb_oof_test.csv', xgb_oof_test, delimiter = ',', fmt = '%.5f')
# # First-level model 2: StackNet
# ## Prepare data for StackNet
# In[18]:
#-------------------------------------------------------------------------------
# First column in train - labels
# First column in test - dummy (indices)
#-------------------------------------------------------------------------------
# Get test index to use as first dummy column in test set for StackNet
ids = test_df['listing_id'].values
# Concat oof and predictions from best model (xgb)
TT_dense = np.c_[TT.toarray(), np.r_[xgb_oof_train, xgb_oof_test]] # (124011, 258)
# Scale
scaler = StandardScaler()
TT_dense = scaler.fit_transform(TT_dense)
# Split
X_train_dense = TT_dense[:r] # (49352, 258)
X_test_dense = TT_dense[r:] # (74659, 258)
# Append target to train
X_train_dense = np.c_[y_train, X_train_dense] # (49352, 259)
# Append id to test
X_test_dense = np.c_[ids, X_test_dense] # (74659, 259)
# Export to txt files
np.savetxt(data_dir + 'train_std.csv', X_train_dense, delimiter = ',', fmt = '%.5f')
np.savetxt(data_dir + 'test_std.csv', X_test_dense, delimiter = ',', fmt = '%.5f')
# ## Run StackNet
# ** * We use dummy 3-level model in file `params.txt` just to get train oof from 2-level model * **
# In[19]:
# Run StackNet and get output
stacknet_log = check_output(['bash', 'run.sh']).decode(sys.stdout.encoding)
# Save output to file
with open(data_dir + 'stacknet_log.txt', 'w') as f:
str_len = f.write(stacknet_log)
# ## Load StackNet OOF
# In[20]:
# Load StackNet oof
stacknet_oof_train = np.loadtxt('stacknet_oof2.csv', delimiter = ',')
stacknet_oof_test = np.loadtxt('stacknet_oof_test2.csv', delimiter = ',')
# # First-level model 3: Extra Trees
# ## Run Extra Trees
# In[21]:
# Inint model
model = ExtraTreesClassifier(random_state = 0, n_jobs = -1, n_estimators = 1000,
criterion = 'entropy', max_depth = None)
# Get oof
et_oof_train, et_oof_test = oof(model, np.c_[X_train_dense[:, 1:], stacknet_oof_train],
y_train, np.c_[X_test_dense[:, 1:], stacknet_oof_test], oof_test = True)
# Export to txt files
np.savetxt(data_dir + 'et_oof_train.csv', et_oof_train, delimiter = ',', fmt = '%.5f')
np.savetxt(data_dir + 'et_oof_test.csv', et_oof_test, delimiter = ',', fmt = '%.5f')
# # Ensemble
# ## Look at oof scores for our models
# In[22]:
# Output oof scores
print('XGB: %.6f' % log_loss(y_train, xgb_oof_train))
print('StackNet: %.6f' % log_loss(y_train, stacknet_oof_train))
print('ET: %.6f' % log_loss(y_train, et_oof_train))
# ## Perform SLSQP optimization with bounds and constraints (9 parameters for each column)
# In[23]:
#-------------------------------------------------------------------------------
# One parameter for each column
#-------------------------------------------------------------------------------
def cost(params):
y_pred = params[:3] * xgb_oof_train + params[3:6] * stacknet_oof_train + params[6:9] * et_oof_train
return log_loss(y_train, y_pred)
def con1(params):
return params[0] + params[3] + params[6] - 1
def con2(params):
return params[1] + params[4] + params[7] - 1
def con3(params):
return params[2] + params[5] + params[8] - 1
# params = [0.33] * 9
# print(cost(params)) # 0.511137
n = 9
init = [0.33] * n
cons = ({'type': 'eq', 'fun': con1},
{'type': 'eq', 'fun': con2},
{'type': 'eq', 'fun': con3})
bounds = [(0, 1)] * n
res = minimize(cost, init, method = 'SLSQP', bounds = bounds, constraints = cons, options = {'maxiter': 100000})
# ## Optimization result
# In[25]:
print(res)
# # Create submission
# In[26]:
params = res['x']
y_pred = params[:3] * xgb_oof_test + params[3:6] * stacknet_oof_test + params[6:9] * et_oof_test
subm_df.loc[:, 'listing_id'] = test_df['listing_id'].values
subm_df.iloc[:, 1:] = y_pred
subm_df.to_csv('./reproduced_submission/reproduced_submission.csv', index = False)
# # Conclusion
# 1. ** * Relatively simple solution with strong result * **
# 2. ** * A lot of fun with cool dataset and competition* **