forked from floridene/deezer_report
-
Notifications
You must be signed in to change notification settings - Fork 1
/
XGBoost_Numeric.rmd
352 lines (266 loc) · 14.6 KB
/
XGBoost_Numeric.rmd
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
## XGBoost Numeric | Binary Logistic | 2017-05-30
##### Score: 0.63073 (Public) 0.63217 (Private) | AUC 0.81 on validation set
#### Use Zeno server (Deezer account) to replicate steps
#### Scroll down to "Prepare XGBoost numeric model" step to run the model with prepared data
#### Pre-processing step
```{r}
#load libraries
library(data.table)
library(jsonlite)
library(caret)
library(xgboost)
library(mice)
library(dplyr)
library(pROC)
library(caTools)
library(tm)
library(qdap)
library(lubridate)
library(httr)
library(Matrix)
#####################################################################################
Feature Engineering
#####################################################################################
#load data (prepared as on 25th May)
load("~/60_data_other_models/Deezer_train_0525.rda")
load("~/60_data_other_models/Deezer_test_0525.rda")
# naming for model
train = DeezerNew_train_0525
test = DeezerNew_test_0525
submission= read.csv("~/60_data_other_models/sample_submission_kaggle.csv")
#label (variable to predict)
y_train = train$is_listened
train_test = rbind(train, test)
#Remove unnecessary columns
train_test$sample_id = NULL
#convert factors to numeric (mandatory for XGBoost model)
train_test %>% mutate_if(is.factor, as.character) -> train_test
train_test %>% mutate_if(is.integer, as.numeric) -> train_test
train_test$user_id = as.numeric(train_test$user_id)
train_test$media_id = as.numeric(train_test$media_id)
train_test$album_id = as.numeric(train_test$album_id)
train_test$context_type = as.numeric(train_test$context_type)
train_test$platform_name = as.numeric(train_test$platform_name)
train_test$platform_family = as.numeric(train_test$platform_family)
train_test$listen_type = as.numeric(train_test$listen_type)
train_test$user_gender = as.numeric(train_test$user_gender)
train_test$artist_id = as.numeric(train_test$artist_id)
train_test$is_listened = as.numeric(train_test$is_listened)
train_test$profile_id = as.numeric(train_test$profile_id)
train_test$hh = as.numeric(train_test$hh)
train_test$wd = as.numeric(train_test$wd)
##################################################
#Pre-processing | Add new features
##################################################
#Bin hours to Timings of the day
train_test$binTime = ifelse(train_test$hh >= 0 & train_test$hh <= 3, "Midnight",
ifelse(train_test$hh >= 4 & train_test$hh <= 5, "EarlyMorning",
ifelse(train_test$hh >= 6 & train_test$hh <= 8, "MidMorning",
ifelse(train_test$hh >= 9 & train_test$hh <= 11, "Morning",
ifelse(train_test$hh >= 12 & train_test$hh <= 16, "Noon",
ifelse(train_test$hh >= 17 & train_test$hh < 21, "Evening", "Night"))))))
#Bin weekends and weekday
train_test$binwkd = ifelse(train_test$wd >= 1 & train_test$wd < 5, "Weekday", "Weekend")
#Bin release years
train_test$binRYear= ifelse(train_test$ryear > 2017, "Outliers",
ifelse(train_test$ryear == 2017, "Recents",
ifelse(train_test$ryear >= 2015 & train_test$ryear <= 2016, "Latest",
ifelse(train_test$ryear >= 2010 & train_test$ryear <= 2014, "FairlyLatest",
ifelse(train_test$ryear >= 2000 & train_test$ryear <= 2009, "Mids1",
ifelse(train_test$ryear >= 1990 & train_test$ryear <= 1999, "Mids2",
ifelse(train_test$ryear >= 1980 & train_test$ryear <= 1989, "Oldies1",
ifelse(train_test$ryear >= 1970 & train_test$ryear <= 1979, "Oldies2",
ifelse(train_test$ryear >= 1950 & train_test$ryear <= 1969, "Oldies3", "VeryOld" )))))))))
# Add new features: join json file with media description (categorical variables)
extra = stream_in(file("~/60_data_other_models/extra_infos.json"))
str(extra)
##################################################
## add language feats (detect lang)
##################################################
library("textcat")
library("rvest")
library("stringr")
#text cleaning (in order to assign unique numeric value)
extra[] = lapply(extra, tolower)
extra[] = lapply(extra, removePunctuation)
extra[] = lapply(extra, stripWhitespace)
extra$songLang = textcat(extra$sng_title)
extra$albLang = textcat(extra$alb_title)
extra$artistLang = textcat(extra$art_name)
#if song lang and alb lang is exact same them grouping them, else considering it as variation
extra$langSngAlb = ifelse(extra$songLang == extra$albLang, extra$albLang, "variation")
extra$langAlbArt = ifelse(extra$albLang == extra$artistLang, extra$artistLang, "variation")
extra$langSngArt = ifelse(extra$songLang == extra$artistLang, extra$artistLang, "variation")
extra$langReg = ifelse(extra$songLang == "french" | extra$songLang == "german" | extra$songLang == "spanish" |
extra$songLang == "swedish" | extra$songLang == "italian" |
extra$songLang == "polish", "TopEuLang", "EnglishOrOther")
str(extra)
extra$media_id = as.numeric(extra$media_id)
# save(extra,file="extra_feats11.rda")
#add to train_test dataset
train_test = left_join(train_test, extra, by = "media_id")
str(train_test)
#Bin by user's age
train_test$binAge = ifelse(train_test$user_age >= 18 & train_test$user_age <= 21 , "yngAds1",
ifelse(train_test$user_age >= 22 & train_test$user_age <= 24 , "yngAds2",
ifelse(train_test$user_age >= 25 & train_test$user_age <= 27 , "yngAds3", "mature")))
tail(train_test$binAge)
str(train_test)
##############################################################
#imputation for NAs
##############################################################
#Imputing lang NAs with french (which is most common)
# hist(train_test$songLang)
as.data.frame(train_test %>% group_by(songLang) %>% summarise(total = n()) %>% arrange(desc(total)))
train_test$songLang[is.na(train_test$songLang)] = "french"
as.data.frame(train_test %>% group_by(albLang) %>% summarise(total = n()) %>% arrange(desc(total)))
train_test$albLang[is.na(train_test$albLang)] = "french"
as.data.frame(train_test %>% group_by(artistLang) %>% summarise(total = n()) %>% arrange(desc(total)))
train_test$artistLang[is.na(train_test$artistLang)] = "english"
as.data.frame(train_test %>% group_by(langAlbArt) %>% summarise(total = n()) %>% arrange(desc(total)))
train_test$langAlbArt[is.na(train_test$langAlbArt)] = "english"
as.data.frame(train_test %>% group_by(langSngArt) %>% summarise(total = n()) %>% arrange(desc(total)))
train_test$langSngArt[is.na(train_test$langSngArt)] = "english"
as.data.frame(train_test %>% group_by(langSngAlb) %>% summarise(total = n()) %>% arrange(desc(total)))
train_test$langSngAlb[is.na(train_test$langSngAlb)] = "english"
as.data.frame(train_test %>% group_by(langReg) %>% summarise(total = n()) %>% arrange(desc(total)))
train_test$langReg[is.na(train_test$langReg)] = "TopEuLang"
as.data.frame(train_test %>% group_by(sng_title) %>% summarise(total = n()) %>% arrange(desc(total)))
train_test$sng_title[is.na(train_test$sng_title)] = "tchikita"
as.data.frame(train_test %>% group_by(alb_title) %>% summarise(total = n()) %>% arrange(desc(total)))
train_test$alb_title[is.na(train_test$alb_title)] = "dans la légende"
# x = as.data.frame(train_test %>% group_by(art_name) %>% summarise(total = n()) %>% arrange(desc(total)))
train_test$art_name[is.na(train_test$art_name)] = "pnl"
sum(is.na(train_test))
summary(train_test)
##############################################################
#Custom relative feats | media duration | track_bpm
##############################################################
summary(train_test$media_duration)
Freq_md = as.data.frame(train_test %>% group_by(media_duration) %>% summarise(total = n()) %>%
arrange(desc(total)))
Freq_md1 = as.data.frame(train_test %>% group_by(media_duration) %>% summarise(total = n()) %>%
arrange(desc(media_duration)))
hist(Freq_md$media_duration)
#Bin by media_duration and user_gender
train_test$md_mf_group = ifelse(train_test$media_duration <= 100 & train_test$user_gender == 0, 1,
ifelse(train_test$media_duration <= 100 & train_test$user_gender == 1, 2,
ifelse(train_test$media_duration >= 101 & train_test$media_duration <= 200 &
train_test$user_gender == 0, 3,
ifelse(train_test$media_duration >= 101 & train_test$media_duration <= 200 &
train_test$user_gender == 1, 4,
ifelse(train_test$media_duration >= 201 & train_test$media_duration <= 300 &
train_test$user_gender == 0, 5,
ifelse(train_test$media_duration >= 201 & train_test$media_duration <= 300 &
train_test$user_gender == 1, 6,
ifelse(train_test$media_duration >= 301 & train_test$media_duration <= 400 &
train_test$user_gender == 0, 7,
ifelse(train_test$media_duration >= 301 & train_test$media_duration <= 400 &
train_test$user_gender == 1, 8,
ifelse(train_test$media_duration >= 401 & train_test$media_duration <= 500 &
train_test$user_gender == 0, 9,
ifelse(train_test$media_duration >= 401 & train_test$media_duration <= 500 &
train_test$user_gender == 1, 10, 11))))))))))
hist(train_test$md_mf_group)
str(train_test)
##############################################
# Save dataset with engieered features
##############################################
# Covert to all numeric data (This is primary requirement of XGboost model)
train_test %>% mutate_if(is.factor, as.character) -> train_test
# label encoding
features= names(train_test)
for (f in features) {
if (class(train_test[[f]])=="character") {
levels <- unique(c(train_test[[f]]))
train_test[[f]] <- as.integer(factor(train_test[[f]], levels=levels))
}
}
#convert into numeric
train_test[] <- lapply(train_test, as.numeric)
str(train_test)
summary(train_test)
#complied rda file with 34 features (ALL NUMERIC)
save(train_test,file="train_test_NUM_20170526_P.rda")
```
##############################################
## Prepare XGBoost numeric model
##############################################
```{r}
#load data (all numeric)
load(~/60_data_other_models/train_test_NUM_20170526_P.rda")
#response variable
str(train_test$is_listened)
train_test = train_test %>% select(-c(user_id, media_id, album_id, artist_id))
#create 3 sets
train = train_test[1:7538916, ]
valid = train_test[7538917:7558834, ]
test = train_test[7558835:7578752, ]
#Convert to matrix
xgb.data.train <- xgb.DMatrix(as.matrix(train[, colnames(train) != "is_listened"]), label = train$is_listened)
xgb.data.valid <- xgb.DMatrix(as.matrix(valid[, colnames(valid) != "is_listened"]), label = valid$is_listened)
xgb.data.test <- xgb.DMatrix(as.matrix(test[, colnames(test) != "is_listened"]), label = test$is_listened)
# Train xgb model (reduced rounds for test purpose)
xgb.model.binLogit <- xgb.train(data = xgb.data.train,
params = list(objective = "binary:logistic",
eta = 0.1,
max.depth = 12,
min_child_weight = 100,
subsample = 0.8,
colsample_bytree = 0.8,
nthread = 4,
eval_metric = "auc"),
watchlist = list(valid = xgb.data.valid),
nrounds = 10,
early_stopping_rounds = 3,
print_every_n = 1)
# Train xgb model (Original)
#xgb.model.binLogit <- xgb.train(data = xgb.data.train,
# params = list(objective = "binary:logistic",
# eta = 0.1,
# max.depth = 12,
# min_child_weight = 100,
# subsample = 0.8,
# colsample_bytree = 0.8,
# nthread = 4,
# eval_metric = "auc"),
# watchlist = list(valid = xgb.data.valid),
# nrounds = 3000,
# early_stopping_rounds = 100,
# print_every_n = 10)
#[388] valid-auc:0.811852
print(xgb.model.binLogit)
print(xgb.model.binLogit$best_score)
xgb.model.binLogit$best_ntreelimit
# save model to binary local file
xgb.save(xgb.model.binLogit, "xgb_woIDs_0.81AUC_NUM_BinLogit_0530")
#Get feature importance
xgb.feature.imp = xgb.importance(model = xgb.model.binLogit)
##########################################################################################
# Make predictions on validation set for ROC curve
xgb.valid.acc = predict(xgb.model.binLogit
, newdata = as.matrix(valid[, colnames(valid) != "is_listened"])
, ntreelimit = xgb.model.binLogit$best_ntreelimit)
auc.xgb.acc_v = roc(valid$is_listened, xgb.valid.acc, plot = TRUE, col = "green")
print(auc.xgb.acc_v)
#0.81
##########################################################################################
# Make predictions on TEST set
xgb.test.acc = predict(xgb.model.binLogit,
newdata = as.matrix(test[, colnames(test) != "is_listened"]),
ntreelimit = xgb.model.binLogit$best_ntreelimit)
## Create submission file
submission <- read.csv("~/R/DSG 17/525NewFeats/sample_submission_kaggle.csv")
submission$is_listened <- xgb.test.acc
write.csv(submission,"Preds_xgb_without_IDs_0.81AUC_NUM_BinLogit_0530.csv",row.names = FALSE)
hist(submission$is_listened, col = "dodgerblue4",
main = "Preds distribution | No Ids | XGB BinLogit | 0.81 AUC | 30 feats | All Numeric")
# Score: 0.63073 (Public) 0.63217 (Private)
```
### Feature importance matrix
* XGB Numeric model with 49 features dated 13th May
#### "Every little improvement in accuracy matters!"
* XGB Numeric model with 33 features dated 18th May
