-
Notifications
You must be signed in to change notification settings - Fork 0
/
settings.R
513 lines (414 loc) · 18.7 KB
/
settings.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
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
# 0. package call ---------------------------------------------------------
# For preprocessing
library(stringr)
library(dplyr)
# For fitting interpretable model
library(glmnet) # LASSO-penalized GLM
library(mboost) # Componentwise Gradient Boosting
# For fitting predictive model
library(caret) # To use confusionMatrix
library(randomForest) # randomForest
library(e1071) # SVM
library(xgboost) # XGBoost
library(mgcv) # Ensemble through GAM
# 1. preprocessing --------------------------------------------------------
logical_match <- function (input_data) {
cleaned <- input_data
# 1. Revise loan amount to zero
# Client whose total loan record equals to 0 will have value of 4
loan_count <- (input_data[, 3] == 0) +
(input_data[, 4] == 0) +
(input_data[, 5] == 0) +
(input_data[, 6] == 0)
# Revise loan amount of such people to zero
cleaned[loan_count == 4, 7] <- 0
cleaned[loan_count == 4, 8] <- 0
# 2. Revise delay ratio to zero
# Client whose max monthly input equals to 0 will have value of TRUE
maxinput_zero <- input_data[, 44] == 0
# Revise their
cleaned[maxinput_zero, 34] <- 0 # total delay ratio
cleaned[maxinput_zero, 35] <- 0 # delay ratio in latest year
cleaned[maxinput_zero, 40] <- 0 # delay ratio of assurance product
cleaned[maxinput_zero, 41] <- 0 # delay ratio of depository product
# to zero
# 3. Revise insurance return to zero
# Client who doesn't have history of request on insurance return will have value of TRUE
request_zero <- input_data[, 51] == 0
cleaned[request_zero, 50] <- 0
return(cleaned)
}
rightmost_cleaner <- function (column) {
# Replacing trivial '1' at the rightmost letter
target <- column
# Replace rightmost letter of element whose rightmost letter is 1 to 0
str_sub(target[str_sub(target, -1, -1) == 1], -1 , -1) <- 0
return(as.integer(target))
}
scaling_column <- function (input_data) {
cleaned <- input_data
# Different company used different unit(SCI : 1000, Hanhwa : 10000)
cleaned <- cleaned %>%
mutate_at(c(7:10, 16, 18, 19, 24), rightmost_cleaner) %>%
mutate_at(c(7:10, 16), list(~ . * 1000)) %>%
mutate_at(c(18, 19, 24), list(~ . * 10000))
return(cleaned)
}
rephrase_date <- function (input_data) {
cleaned <- input_data
# TEL_CNTT_QTR : YYYY/Q(the year and the quarter that the customer enrolled telecom service)
# 20162(second quarter of 2016) was the latest. Thus each value should be rephrased to:
# 20162 -> 0, 20161 -> 1, 20154 -> 2, 20153 -> 3, 20152 -> 4, 20151 -> 5, 20144 -> 6, ...
#
# To obtain such mapping, first 20162 is subtracted from each value.
# Then with its absoulute value, following mapping rule is applied:
# (1) If first digit is less than 7, save that digit as it is;
# if it's over 7, subtract 6 from that digit and save that value.
# (2) Next, divide that value by 10 and discard the remaining.
# (3) Then add (that quotient) * 4 to obtain mapping
#
# Ex) map 20144 to 6
# (1) |20144 - 20162| = 18 (First digit is 8; therefore save 8 - 6 = 2)
# (2) 18 / 10 = 1.8 -> 1 (discard the remaining)
# (3) Add 1 * 4 = 4 to the value(i.e. 2 + 4 = 6)
# initialize
dist_year <- abs(input_data[, 62] - 20162)
# (1)
first_digit <- as.numeric(str_sub(dist_year, -1, -1))
first_digit[first_digit > 7] <- first_digit[first_digit > 7] - 6
# (2)
dist_year <- floor(dist_year / 10)
# Revise data as planned
cleaned[, 62] <- first_digit + (dist_year * 4)
# MIN_CNTT_DATE : YYYY/MM(year and month that Hanhwa lend the money to the client first time)
# Same idea can be applied, but 90% of data already has value of 0
# (i.e. 90% of the customer didn't borrowed money from Hanhwa;
# showing the lackness of lending experience toward middle-credit customers)
year <- str_sub(input_data[, 26], 1, 4)
month <- str_sub(input_data[, 26], 5, 6)
month[month == ""] <- "0"
distance <- 12 * (2016 - as.integer(year)) + (6 - as.integer(month))
distance[distance > 100] <- 0
cleaned[, 26] <- distance
return(cleaned)
}
trivials <- function (input_data) {
cleaned <- input_data
# latest_binary
latest_binary <- rep(1, nrow(input_data))
latest_binary[input_data[, 22] == 0 | is.null(input_data[, 22])] <- 0
cleaned[, 22] <- as.factor(latest_binary)
rm('latest_binary')
# revise interval into numeric value by using its median
delay_ratio_latest <- input_data[, 35]
cleaner <- rep(0, nrow(cleaned))
cleaner[delay_ratio_latest=='10미만'] <- 5
cleaner[delay_ratio_latest=='20미만'] <- 15
cleaner[delay_ratio_latest=='30미만'] <- 25
cleaner[delay_ratio_latest=='40미만'] <- 35
cleaner[delay_ratio_latest=='50미만'] <- 45
cleaner[delay_ratio_latest=='60미만'] <- 55
cleaner[delay_ratio_latest=='90미만'] <- 75
cleaner[delay_ratio_latest=='90이상'] <- 95
cleaned[, 35] <- cleaner
rm('cleaner', 'delay_ratio_latest')
# Making masked value to zero
age <- as.character(input_data[, 53])
age[age == '*'] <- 0
cleaned[, 53] <- as.integer(age)
rm('age')
return(cleaned)
}
# 1-1. derived variable -----------------------------------------------------
# 1-1-1. By intuition -------------------------------------------------------
add_intuitive <- function (input_data) {
added <- input_data
# (1) lately_delayed
lately_delayed <- rep(0, nrow(input_data))
lately_delayed[input_data$CRMM_OVDU_AMT != 0 | input_data$CRLN_30OVDU_RATE != 0] <- 1
added$lately_delayed <- as.factor(lately_delayed)
# (2) middle_loan_ratio
nan_handler <- input_data$TOT_LNIF_AMT == 0
middle_loan_ratio <- input_data$CPT_LNIF_AMT / input_data$TOT_LNIF_AMT
middle_loan_ratio[nan_handler] <- 0
# To eliminate NaN created by division by zero
added$middle_loan_ratio <- middle_loan_ratio
# (3) amount_per_gurantee
nan_handler <- input_data$CB_GUIF_CNT == 0
amount_per_gurantee <- input_data$CB_GUIF_AMT/input_data$CB_GUIF_CNT
amount_per_gurantee[nan_handler] <- 0
# To eliminate NaN created by division by zero
added$amount_per_gurantee <- amount_per_gurantee
# (4) amount_per_bankloan
nan_handler <- input_data$BNK_LNIF_CNT == 0
amount_per_bankloan <- input_data$BNK_LNIF_AMT / input_data$BNK_LNIF_CNT
amount_per_bankloan[nan_handler] <- 0
# To eliminate NaN created by division by zero
added$amount_per_bankloan <- amount_per_bankloan
return(added)
}
# 1-1-2. By Chi-square partition --------------------------------------------
revalue_column_conttable <- function (input_data, target, nlevels_groupA) {
if (is.character(target)) {
target_index <- which(names(input_data) == target)
} else {
target_index <- target
} # whether target is index of variable or variable's name,
# target_index will always contain index of the variable.
levels_groupA <- levels(factor(input_data[, target_index]))[1:nlevels_groupA]
# levels_groupA is a set of values that will be encoded as group A.
revalue_result <- rep("B", nrow(input_data))
revalue_result[input_data[, target_index] %in% levels_groupA] <- "A"
# if value of a cell is element of levels_groupA, encode it as 'A'
return(revalue_result)
}
ordinal_signif_conttable <- function (input_data, index) {
index_level <- as.numeric(levels(factor(input_data[, index])))
seperator <- numeric(length(index_level) - 1)
chisq <- numeric(length(index_level) - 1)
for(i in 1:(length(index_level) - 1)){
# (1) Bisect the variable
grouped <- revalue_column_conttable(input_data, index, i)
# (2) Build contingency table
critical_table <- table(input_data$TARGET, grouped)
# (3) Calculate Chi-square statistic
a <- as.numeric(critical_table[1, 1])
b <- as.numeric(critical_table[1, 2])
c <- as.numeric(critical_table[2, 1])
d <- as.numeric(critical_table[2, 2])
n <- sum(critical_table)
seperator[i] <- index_level[i]
chisq[i] <- n * ((a * d - b * c) ^ 2) / ((a + b) * (c + d) * (a + c) * (b + d))
}
return(data.frame(sep = seperator, chisquare = chisq))
}
compound_2var <- function(input_data, index1, index2, nlevels_A1, nlevels_A2){
# 1. Bisect two variables
revalue1 <- revalue_column_conttable(input_data, index1, nlevels_A1)
revalue2 <- revalue_column_conttable(input_data, index2, nlevels_A2)
# 2. Label the case into one of four
# 2-1. Initialize every label to zero
complected <- rep(0, nrow(input_data))
# 2-2. Filter the values whose first variable belongs to group B
# and label them as either 0 or 2
complected[revalue1 == 'B'] <- 2
# 2-3. Filter the values whose second variable belongs to B and add 1
# (then they will be labeled as either 1 or 3;
# 1 if first = A and second = B, 3 if first = B and second = B)
# and replace this value to complete labelling
complected[revalue2 == 'B'] <- complected[revalue2 == 'B'] + 1
return(complected)
}
compound3456 <- function (input_data) {
# Store bisected result of 4 variables
subsetted_factored <- as.data.frame(
cbind(revalue_column_conttable(input_data, 3, 1),
revalue_column_conttable(input_data, 4, 1),
revalue_column_conttable(input_data, 5, 2),
revalue_column_conttable(input_data, 6, 2))
)
# Label 16 different cases
# ; using the same labeling technique in compound_2var function
category_3456 <- rep(0, nrow(subsetted_factored))
category_3456[subsetted_factored[, 1] == 'B'] <- category_3456[subsetted_factored[, 1] == 'B'] + 1
category_3456[subsetted_factored[, 2] == 'B'] <- category_3456[subsetted_factored[, 2] == 'B'] + 2
category_3456[subsetted_factored[, 3] == 'B'] <- category_3456[subsetted_factored[, 3] == 'B'] + 4
category_3456[subsetted_factored[, 4] == 'B'] <- category_3456[subsetted_factored[, 4] == 'B'] + 8
# compound some features
# ; as in phonefee_delaytrend making procedure
category_3456[category_3456 == 0] <- 'A'
category_3456[category_3456 == 1] <- 'A'
category_3456[category_3456 == 2] <- 'B'
category_3456[category_3456 == 3] <- 'A'
category_3456[category_3456 == 4] <- 'B'
category_3456[category_3456 == 5] <- 'B'
category_3456[category_3456 == 6] <- 'B'
category_3456[category_3456 == 7] <- 'B'
category_3456[category_3456 == 8] <- 'C'
category_3456[category_3456 == 9] <- 'B'
category_3456[category_3456 == 10] <- 'B'
category_3456[category_3456 == 11] <- 'B'
category_3456[category_3456 == 12] <- 'C'
category_3456[category_3456 == 13] <- 'C'
category_3456[category_3456 == 14] <- 'C'
category_3456[category_3456 == 15] <- 'C'
return(category_3456)
}
add_chisquare <- function (input_data) {
loaner_effect1 <- compound3456(input_data)
credit_effect <- compound_2var(input_data, 7, 9, 8, 5)
credit_effect[credit_effect == 0] <- 'A'
credit_effect[credit_effect == 1] <- 'B'
credit_effect[credit_effect == 2] <- 'C'
credit_effect[credit_effect == 3] <- 'D'
loaner_effect2 <- compound_2var(input_data, 11, 12, 2, 2)
loaner_effect2[loaner_effect2 == 0] <- 'A'
loaner_effect2[loaner_effect2 == 1] <- 'B'
loaner_effect2[loaner_effect2 == 2] <- 'C'
loaner_effect2[loaner_effect2 == 3] <- 'D'
periodic_failure <- compound_2var(input_data, 49, 52, 1, 1)
periodic_failure[periodic_failure == 1] <- 'A'
periodic_failure[periodic_failure == 0] <- 'B'
periodic_failure[periodic_failure == 3] <- 'C'
periodic_failure[periodic_failure == 2] <- 'D'
phonefee_delaytrend <- compound_2var(input_data, 64, 66, 3, 4)
phonefee_delaytrend[phonefee_delaytrend == 0] <- 'A'
phonefee_delaytrend[phonefee_delaytrend == 1 | phonefee_delaytrend == 2] <- 'B'
phonefee_delaytrend[phonefee_delaytrend == 3] <- 'C'
# Add generated variables
added <- input_data
added$loaner_effect1 <- as.factor(loaner_effect1)
added$credit_effect <- as.factor(credit_effect)
added$loaner_effect2 <- as.factor(loaner_effect2)
added$periodic_failure <- as.factor(periodic_failure)
added$phonefee_delaytrend <- as.factor(phonefee_delaytrend)
return(added)
}
# 2. Fit interpretable model ----------------------------------------------
# 2-1. Significance of Derived variables ----------------------------------
resampler <- function (input_data, tr_rate) {
original_data <- input_data
# 1. Save indices of zero and one proportional to the rate of training data(tr_rate)
# 1-1. index_zero : Randomly select (tr_rate * 100)% of zeros
index_zero <- sample(which(original_data$TARGET == 0),
round(length(which(original_data$TARGET == 0)) * tr_rate, 0),
replace = F)
# 1-2. index_one : Randomly select (tr_rate * 100)% of ones
index_one <- sample(which(original_data$TARGET == 1),
round(length(which(original_data$TARGET == 1)) * tr_rate, 0),
replace = F)
# 2. Save training, test data into the global scope
train_data <<- rbind(original_data[index_zero, ], original_data[index_one, ])
test_data <<- original_data[-as.numeric(rownames(train_data)), ]
}
find_best_cutoff <- function (testdata, cutoff, fitted_glmnet) {
# Subroutine for finding best cutoff point
# 1. Obtain a vector of prediction on test data as a probability(type = 'response')
lambda <- fitted_glmnet$lambda.1se
predicted <- predict(fitted_glmnet,
model.matrix(CUST_ID + TARGET ~ ., data = testdata),
s = lambda,
type = 'response')
# 2. Create a vector to store f-measure values
result <- vector('numeric', length(cutoff))
# 3. Begin recording
for(i in 1:length(cutoff)){
# If probability > cutoff, record it as 1
lasso_result <- rep(0, nrow(testdata))
lasso_result[predicted >= cutoff[i]] <- 1
# calculate F1 score
critical_table <- table(true = testdata$TARGET,
pred = lasso_result)
precision <- critical_table[2, 2] / (critical_table[1, 2] + critical_table[2, 2])
recall <- critical_table[2, 2] / (critical_table[2, 1] + critical_table[2, 2])
F1_score <- 2 * precision * recall / (precision + recall)
# record cutoff point and corresponding F1_score
result <- F1_score
}
# 4. Return the result
return(result)
}
performance_tester_LASSO <- function (training, test, coef_return = TRUE) {
# This function returns the test-F1 score calculated by the best set of estimators for given combination of data.
# Optionally returns corresponding selected coefficients.
#
# 1. Applying Cross-Validation to find best coefficient estimate
#
# Since this is highly unbalanced data, measuring the fit simply by misclassification error
# (i.e. type.measure = 'class') seemed inappropriate. Therefore, AUC is used as a measure of the fit.
logistic_fit <- cv.glmnet(model.matrix(CUST_ID + TARGET ~ ., data = training), training$TARGET,
alpha = 1, family = 'binomial', type.measure = 'auc')
# 2. Obtain the list of F1 scores
if (coef_return) {
return(list(fmeasure = find_best_cutoff(test, cutoff_points, logistic_fit),
coef = data.frame(
names = rownames(coef(logistic_fit, s = logistic_fit$lambda.1se)),
values = as.numeric(coef(logistic_fit, s = logistic_fit$lambda.1se))
)))
} else {
return(fmeasure = find_best_cutoff(test, cutoff_points, logistic_fit))
}
}
# 2-2. Justification on Boosting method -------------------------------------
performance_tester_Boosting <- function (training, test, coef_return = TRUE) {
# This function returns the test-F1 score calculated by the best set of estimators for given combination of data.
# Optionally returns corresponding selected coefficients.
#
# 1. Obtain fitted logistic regression using Componentwise Boosting method
logistic_fit <- glmboost(TARGET ~ .,
data = training,
family = Binomial(type='glm', link='logit'),
control=boost_control(mstop=10000))
# 2. Obtain the list of F1 scores
result <- matrix(0, nrow = 2, ncol = length(cutoff_points))
for (i in 1:length(cutoff_points)) {
gbtable <- table(data$TARGET,
ifelse(predict(gbobject, newdata = test, type='response') > cutoff[i], 1, 0))
a <- gbtable[2,2] / (gbtable[1, 2] + gbtable[2, 2])
b <- gbtable[2,2] / (gbtable[2, 1] + gbtable[2, 2])
result[1, i] <- cutoff[i]
result[2, i] <- 2 * a * b / (a + b)
}
highest <- max(result[, 2])
if (coef_return) {
return(list(fmeasure = highest,
cutoff = result[which(result[, 2] == highest), 1],
coefficient = coef(logistic_fit)
))
} else {
return(highest)
}
}
# 3. Fit predictive model -------------------------------------------------
dummy_column_maker <- function (column) {
# input : selected column of the data
# output : dummy-processed result of that column
level_names <- levels(factor(column))
nlevel <- length(level_names)
storage <- matrix(rep(0, length(column) * nlevel), ncol = nlevel)
# Create dummy column for each level
for (i in 1:nlevel) storage[column == level_names[i], i] <- 1
# Return the result as data frame
result <- as.data.frame(storage)
names(result) <- paste0('_', level)
return(result)
}
dummy_process <- function(input_data, colname){
colnum <- which(names(input_data) == colname)
if (colnum == length(input_data)) {
# If the last variable of the data is factor variable, dummy-process that last column and return it as NewData.
NewData <- cbind(input_data[, 1:(colnum - 1)], dummy_column_maker(input_data[, colnum]))
} else {
NewData <- cbind(input_data[, 1:(colnum - 1)],
dummy_column_maker(input_data[,colnum]),
input_data[, (colnum + 1):length(input_data)])
}
return(NewData)
}
dummy_processed_data = function(data){
# 1. Store the name of the column whose value is in factor type
FacCol <- c()
for (i in 1:length(data)){
if (class(data[,i]) == 'factor') FacCol <- c(FacCol, names(data)[i])
}
# 2. Conduct dummy_process to that data
processed_data <- dummy_process(data, FacCol[1])
for (i in 2:length(FacCol)) processed_data <- dummy_process(processed_data, FacCol[i])
return(processed_data)
}
getF1score <- function (cmat = NULL, test = NA, pred = NA) {
if (is.na(test)) {
# If entered confusionMatrix(i.e. in case of RF, SVM)
precision <- cmat$table[2, 2] / (cmat$table[1, 2] + cmat$table[2, 2])
recall <- cmat$table[2, 2] / (cmat$table[2, 1] + cmat$table[2, 2])
f1score <- 2 * precision * recall / (precision + recall)
return(f1score)
} else {
preds <- as.numeric(pred > 0.2)
labels <- test$TARGET
precision <- sum(preds == 1 & labels == 1) / sum(labels == 1)
recall <- sum(preds == 1 & labels == 1) / sum(preds == 1)
f1score <- (2 * precision * recall) / (precision + recall)
return(f1score)
}
}