-
Notifications
You must be signed in to change notification settings - Fork 15
/
model_xgb.R
313 lines (240 loc) · 12.1 KB
/
model_xgb.R
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
#### Santander Product Recommendation on Kaggle
#### Team 'SRRRK' (Sudalai Raj Kumar and Rohan Rao)
#### 11th Rank
## setting file paths and seed (edit the paths before running)
path_train <- "train.csv"
path_test <- "test.csv"
path_preds <- "preds.csv"
## put your favourite number as seed
seed <- 123
set.seed(seed)
## loading libraries
library(data.table)
library(xgboost)
## loading raw data
train <- fread(path_train, showProgress = T)
test <- fread(path_test, showProgress = T)
## removing five products
train[, ind_ahor_fin_ult1 := NULL]
train[, ind_aval_fin_ult1 := NULL]
train[, ind_deco_fin_ult1 := NULL]
train[, ind_deme_fin_ult1 := NULL]
train[, ind_viv_fin_ult1 := NULL]
## extracting train data of each product and rbinding them together with single multiclass label
i <- 0
target_cols <- names(train)[which(regexpr("ult1", names(train)) > 0)]
for (target_col in target_cols)
{
i <- i + 1
S <- paste0("train", i, " <- train[", target_col, " > 0]")
eval(parse(text = S))
}
rm(train)
gc()
for (i in 1:19)
{
S1 <- paste0("train", i, " <- train", i, "[, !target_cols, with = F]")
eval(parse(text = S1))
S2 <- paste0("train", i, "[, target := ", i-1, "]")
eval(parse(text = S2))
}
X_train <- rbind(train1, train2, train3, train4, train5, train6, train7, train8, train9, train10,
train11, train12, train13, train14, train15, train16, train17, train18, train19)
rm(train1, train2, train3, train4, train5, train6, train7, train8, train9, train10,
train11, train12, train13, train14, train15, train16, train17, train18, train19)
gc()
## rbinding train and test data
X_panel <- rbind(X_train, test, use.names = T, fill = T)
## adding corresponding numeric months (1-18) to fecha_dato
X_panel[, month := as.numeric(as.factor(fecha_dato))]
## creating user-product matrix
X_user_target <- dcast(X_panel[!is.na(target)], ncodpers + month ~ target, length, value.var = "target", fill = 0)
## creating product lag-variables of order-12 and merging with data
X_user_target[, month := month + 1]
setnames(X_user_target,
c("0", "1", "2", "3", "4", "5", "6", "7", "8", "9", "10", "11", "12", "13", "14", "15",
"16", "17", "18"),
c("prev_0", "prev_1", "prev_2", "prev_3", "prev_4", "prev_5", "prev_6", "prev_7",
"prev_8", "prev_9", "prev_10", "prev_11", "prev_12", "prev_13", "prev_14", "prev_15",
"prev_16", "prev_17", "prev_18"))
X_panel <- merge(X_panel, X_user_target, all.x = T, by = c("ncodpers", "month"))
for (i in 2:12)
{
X_user_target[, month := month + 1]
lag_cols <- names(X_user_target)[which(regexpr("prev", names(X_user_target)) > 0)]
setnames(X_user_target, lag_cols, paste0("prev_", lag_cols))
X_panel <- merge(X_panel, X_user_target, all.x = T, by = c("ncodpers", "month"))
}
X_panel[is.na(X_panel)] <- 0
## calculating historic average of product lag-variables with 1, 2, 3, 4, 5, 6, 9, 12 months
for (i in 0:18)
{
S1 <- paste0("X_panel[, prev2_", i, " := (prev_", i, " + prev_prev_", i, ") / 2]")
eval(parse(text = S1))
S2 <- paste0("X_panel[, prev3_", i, " := (prev_", i, " + prev_prev_", i, " + prev_prev_prev_", i, ") / 3]")
eval(parse(text = S2))
S3 <- paste0("X_panel[, prev4_", i, " := (prev_", i, " + prev_prev_", i, " + prev_prev_prev_", i, " + prev_prev_prev_prev_", i, ") / 4]")
eval(parse(text = S3))
S4 <- paste0("X_panel[, prev5_", i, " := (prev_", i, " + prev_prev_", i, " + prev_prev_prev_", i, " + prev_prev_prev_prev_", i, " + prev_prev_prev_prev_prev_", i, ") / 5]")
eval(parse(text = S4))
S5 <- paste0("X_panel[, prev6_", i, " := (prev_", i, " + prev_prev_", i, " + prev_prev_prev_", i, " + prev_prev_prev_prev_", i, " + prev_prev_prev_prev_prev_", i, " + prev_prev_prev_prev_prev_prev_", i, ") / 6]")
eval(parse(text = S5))
S6 <- paste0("X_panel[, prev9_", i, " := (prev_", i, " + prev_prev_", i, " + prev_prev_prev_", i, " + prev_prev_prev_prev_", i, " + prev_prev_prev_prev_prev_", i, " + prev_prev_prev_prev_prev_prev_", i, "
+ prev_prev_prev_prev_prev_prev_prev_", i, " + prev_prev_prev_prev_prev_prev_prev_prev_", i, " + prev_prev_prev_prev_prev_prev_prev_prev_prev_", i, ") / 9]")
eval(parse(text = S6))
S6 <- paste0("X_panel[, prev12_", i, " := (prev_", i, " + prev_prev_", i, " + prev_prev_prev_", i, " + prev_prev_prev_prev_", i, " + prev_prev_prev_prev_prev_", i, " + prev_prev_prev_prev_prev_prev_", i, "
+ prev_prev_prev_prev_prev_prev_prev_", i, " + prev_prev_prev_prev_prev_prev_prev_prev_", i, " + prev_prev_prev_prev_prev_prev_prev_prev_prev_", i, "
+ prev_prev_prev_prev_prev_prev_prev_prev_prev_prev_", i, " + prev_prev_prev_prev_prev_prev_prev_prev_prev_prev_prev_", i, " + prev_prev_prev_prev_prev_prev_prev_prev_prev_prev_prev_prev_", i, ") / 12]")
eval(parse(text = S6))
S7 <- paste0("X_panel[, ':='(prev_prev_", i, " = NULL, prev_prev_prev_", i, " = NULL, prev_prev_prev_prev_", i, " = NULL, prev_prev_prev_prev_prev_", i, " = NULL, prev_prev_prev_prev_prev_prev_", i, " = NULL,
prev_prev_prev_prev_prev_prev_prev_", i, " = NULL, prev_prev_prev_prev_prev_prev_prev_prev_", i, " = NULL, prev_prev_prev_prev_prev_prev_prev_prev_prev_", i, " = NULL,
prev_prev_prev_prev_prev_prev_prev_prev_prev_prev_", i, " = NULL, prev_prev_prev_prev_prev_prev_prev_prev_prev_prev_prev_", i, " = NULL, prev_prev_prev_prev_prev_prev_prev_prev_prev_prev_prev_prev_", i, " = NULL)]")
eval(parse(text = S7))
}
## cleaning raw features
X_panel[, ":="(ind_empleado = as.numeric(as.factor(ind_empleado)),
pais_residencia = as.numeric(as.factor(pais_residencia)),
sexo = as.numeric(as.factor(sexo)),
year_joining = year(as.Date(fecha_alta)),
month_joining = month(as.Date(fecha_alta)),
fecha_alta = as.numeric(as.Date(fecha_alta) - as.Date("2016-05-31")),
ult_fec_cli_1t = ifelse(ult_fec_cli_1t == "", 0, 1),
indrel_1mes = as.numeric(as.factor(indrel_1mes)),
tiprel_1mes = as.numeric(as.factor(tiprel_1mes)),
indresi = as.numeric(as.factor(indresi)),
indext = as.numeric(as.factor(indext)),
conyuemp = as.numeric(as.factor(conyuemp)),
canal_entrada = as.numeric(as.factor(canal_entrada)),
indfall = as.numeric(as.factor(indfall)),
tipodom = NULL,
cod_prov = as.numeric(as.factor(cod_prov)),
nomprov = NULL,
segmento = as.numeric(as.factor(segmento)))]
## calculating product count of previous month
X_panel[, count_prev1_products := (prev_0 + prev_1 + prev_2 + prev_3 + prev_4 + prev_5 + prev_6
+ prev_7 + prev_8 + prev_9 + prev_10 + prev_11 + prev_12
+ prev_13 + prev_14 + prev_15 + prev_16 + prev_17 + prev_18)]
## label encoding binary string of previous month's products (in order of popularity)
X_panel[, prev_products := as.numeric(as.factor(paste0(prev_0, prev_18, prev_5, prev_2, prev_8,
prev_17, prev_16, prev_13, prev_14,
prev_15, prev_6, prev_7, prev_9, prev_4,
prev_3, prev_11, prev_10, prev_12, prev_1)))]
## creating train and test data for June-15 (seasonality) and May-16 (trend) models
X_train_1 <- X_panel[fecha_dato %in% c("2015-06-28")]
X_train_2 <- X_panel[fecha_dato %in% c("2016-05-28")]
X_test_1 <- X_panel[fecha_dato %in% c("2016-06-28")]
X_test_2 <- X_panel[fecha_dato %in% c("2016-06-28")]
X_test_order <- X_test_1$ncodpers
## creating binary flag for new products, test data will always have 1 since we need to predict new products
X_train_1$flag_new <- 0
X_train_2$flag_new <- 0
X_test_1$flag_new <- 1
X_test_2$flag_new <- 1
for (i in 0:18)
{
S1 <- paste0("X_train_1$flag_new[X_train_1$prev_", i, " == 0 & X_train_1$target == ", i, "] <- 1")
eval(parse(text = S1))
S2 <- paste0("X_train_2$flag_new[X_train_2$prev_", i, " == 0 & X_train_2$target == ", i, "] <- 1")
eval(parse(text = S2))
}
## removing lag6, lag9, lag12 variables from seasonality model
for (col in names(X_train_1)[which(regexpr("prev6", names(X_train_1)) > 0 | regexpr("prev9", names(X_train_1)) > 0 | regexpr("prev12", names(X_train_1)) > 0)])
{
S1 <- paste0("X_train_1[, ", col, " := NULL]")
eval(parse(text = S1))
S2 <- paste0("X_test_1[, ", col, " := NULL]")
eval(parse(text = S2))
}
## extracting labels
X_target_1 <- X_train_1$target
X_target_2 <- X_train_2$target
## removing redundant columns
X_train_1[, ":="(fecha_dato = NULL, ncodpers = NULL, month = NULL, target = NULL)]
X_train_2[, ":="(fecha_dato = NULL, ncodpers = NULL, month = NULL, target = NULL)]
X_test_1[, ":="(fecha_dato = NULL, ncodpers = NULL, month = NULL, target = NULL)]
X_test_2[, ":="(fecha_dato = NULL, ncodpers = NULL, month = NULL, target = NULL)]
## creating xgb.DMatrix
xgtrain1 <- xgb.DMatrix(as.matrix(X_train_1), label = X_target_1, missing = NA)
xgtrain2 <- xgb.DMatrix(as.matrix(X_train_2), label = X_target_2, missing = NA)
xgtest1 <- xgb.DMatrix(as.matrix(X_test_1), missing = NA)
xgtest2 <- xgb.DMatrix(as.matrix(X_test_2), missing = NA)
## xgboost parameters
params <- list()
params$objective <- "multi:softprob"
params$num_class <- 19
params$eta <- 0.1
params$max_depth <- 5
params$subsample <- 0.8
params$colsample_bytree <- 0.8
params$min_child_weight <- 3
params$eval_metric <- "mlogloss"
## xgboost training
model_xgb_1 <- xgb.train(params = params, xgtrain1, nrounds = 140, nthread = -1)
model_xgb_2 <- xgb.train(params = params, xgtrain2, nrounds = 140, nthread = -1)
## xgboost predictions
pred_1 <- predict(model_xgb_1, xgtest1)
pred_2 <- predict(model_xgb_2, xgtest2)
## converting predictions into 19 product columns
pred_matrix_1 <- data.table(matrix(pred_1, ncol = 19, byrow = T))
pred_matrix_2 <- data.table(matrix(pred_2, ncol = 19, byrow = T))
## cco weights
pred_matrix_1[,1] <- 0.9 * pred_matrix_1[,1]
pred_matrix_2[,1] <- 0.1 * pred_matrix_2[,1]
## cder weights
pred_matrix_1[,2] <- 0.5 * pred_matrix_1[,2]
pred_matrix_2[,2] <- 0.5 * pred_matrix_2[,2]
## cno weights
pred_matrix_1[,3] <- 0.3 * pred_matrix_1[,3]
pred_matrix_2[,3] <- 0.7 * pred_matrix_2[,3]
## ctju weights
pred_matrix_1[,4] <- 0.4 * pred_matrix_1[,4]
pred_matrix_2[,4] <- 0.6 * pred_matrix_2[,4]
## ctma weights
pred_matrix_1[,5] <- 0.3 * pred_matrix_1[,5]
pred_matrix_2[,5] <- 0.7 * pred_matrix_2[,5]
## ctop weights
pred_matrix_1[,6] <- 0.5 * pred_matrix_1[,6]
pred_matrix_2[,6] <- 0.5 * pred_matrix_2[,6]
## ctpp weights
pred_matrix_1[,7] <- 0.4 * pred_matrix_1[,7]
pred_matrix_2[,7] <- 0.6 * pred_matrix_2[,7]
## dela weights
pred_matrix_1[,8] <- 0.1 * pred_matrix_1[,8]
pred_matrix_2[,8] <- 0.9 * pred_matrix_2[,8]
## ecue weights
pred_matrix_1[,9] <- 0.3 * pred_matrix_1[,9]
pred_matrix_2[,9] <- 0.7 * pred_matrix_2[,9]
## fond weights
pred_matrix_1[,10] <- 0.1 * pred_matrix_1[,10]
pred_matrix_2[,10] <- 0.9 * pred_matrix_2[,10]
## hip weights
pred_matrix_1[,11] <- 0.5 * pred_matrix_1[,11]
pred_matrix_2[,11] <- 0.5 * pred_matrix_2[,11]
## plan weights
pred_matrix_1[,12] <- 0.5 * pred_matrix_1[,12]
pred_matrix_2[,12] <- 0.5 * pred_matrix_2[,12]
## pres weights
pred_matrix_1[,13] <- 0.5 * pred_matrix_1[,13]
pred_matrix_2[,13] <- 0.5 * pred_matrix_2[,13]
## reca weights
pred_matrix_1[,14] <- 0.9 * pred_matrix_1[,14]
pred_matrix_2[,14] <- 0.1 * pred_matrix_2[,14]
## tjcr weights
pred_matrix_1[,15] <- 0.5 * pred_matrix_1[,15]
pred_matrix_2[,15] <- 0.5 * pred_matrix_2[,15]
## valo weights
pred_matrix_1[,16] <- 0.6 * pred_matrix_1[,16]
pred_matrix_2[,16] <- 0.4 * pred_matrix_2[,16]
## nomina weights
pred_matrix_1[,17] <- 0.8 * pred_matrix_1[,17]
pred_matrix_2[,17] <- 0.2 * pred_matrix_2[,17]
## nom_pens weights
pred_matrix_1[,18] <- 0.8 * pred_matrix_1[,18]
pred_matrix_2[,18] <- 0.2 * pred_matrix_2[,18]
## recibo weights
pred_matrix_1[,19] <- 0.4 * pred_matrix_1[,19]
pred_matrix_2[,19] <- 0.6 * pred_matrix_2[,19]
## saving preds to csv
pred_matrix <- pred_matrix_1 + pred_matrix_2
pred_matrix[, ncodpers := X_test_order]
fwrite(pred_matrix, path_preds)