use_grnn.R

df <- readRDS("df.rds") 
source("mob.R")
source("grnnet.R")

# PRE-PROCESS THE DATA WITH MOB PACKAGE
bin_out <- batch_bin(df, 3)
bin_out$BinSum[order(-bin_out$BinSum$iv), ]
#            var nbin unique miss min  median     max      ks     iv
#   bureau_score   34    315  315 443   692.5     848 35.2651 0.8357
#   tot_rev_line   20   3617  477   0 10573.0  205395 26.8943 0.4442
#  age_oldest_tr   25    460  216   1   137.0     588 20.3646 0.2714
#      tot_derog    7     29  213   0     0.0      32 20.0442 0.2599
#            ltv   17    145    1   0   100.0     176 16.8807 0.1911
#       rev_util   12    101    0   0    30.0     100 16.9615 0.1635
#         tot_tr   15     67  213   0    16.0      77 17.3002 0.1425
#   tot_rev_debt    8   3880  477   0  3009.5   96260  8.8722 0.0847
#     tot_rev_tr    4     21  636   0     3.0      24  9.0779 0.0789
#     tot_income   17   1639    5   0  3400.0 8147167 10.3386 0.0775
#    tot_open_tr    7     26 1416   0     5.0      26  6.8695 0.0282

# PERFORMAN WOE TRANSFORMATIONS
df_woe <- batch_woe(df, bin_out$BinLst)

# PROCESS AND STANDARDIZE THE DATA WITH ZERO MEAN AND UNITY VARIANCE
Y <- df$bad
X <- scale(df_woe$df[, -1])
Reduce(rbind, Map(function(c) data.frame(var = colnames(X)[c], mean = mean(X[, c]), variance = var(X[, c])), seq(dim(X)[2])))
#                 var          mean variance
#1      woe.tot_derog  2.234331e-16        1
#2         woe.tot_tr -2.439238e-15        1
#3  woe.age_oldest_tr -2.502177e-15        1
#4    woe.tot_open_tr -2.088444e-16        1
#5     woe.tot_rev_tr -4.930136e-15        1
#6   woe.tot_rev_debt -2.174607e-16        1
#7   woe.tot_rev_line -8.589630e-16        1
#8       woe.rev_util -8.649849e-15        1
#9   woe.bureau_score  1.439904e-15        1
#10           woe.ltv  3.723332e-15        1
#11    woe.tot_income  5.559240e-16        1

# INITIATE A GRNN OBJECT 
net1 <- grnn.fit(x = X, y = Y)
# CROSS-VALIDATION TO CHOOSE THE OPTIONAL SMOOTH PARAMETER
S <- gen_sobol(min = 0.5, max = 1.5, n = 10, seed = 2019)
cv <- grnn.cv_auc(net = net1, sigmas = S, nfolds = 5)
# $test
#       sigma       auc
#1  1.4066449 0.7543912
#2  0.6205723 0.7303415
#3  1.0710133 0.7553075
#4  0.6764866 0.7378430
#5  1.1322939 0.7553664
#6  0.8402438 0.7507192
#7  1.3590402 0.7546164
#8  1.3031974 0.7548670
#9  0.7555905 0.7455457
#10 1.2174429 0.7552097
# $best
#     sigma       auc
#5 1.132294 0.7553664

# REFIT A GRNN WITH THE OPTIMAL PARAMETER VALUE
net2 <- grnn.fit(x = X, y = Y, sigma = cv$best$sigma)
net2.pred <- grnn.parpred(net2, X)

# BENCHMARK MODEL PERFORMANCE
MLmetrics::KS_Stat(y_pred = net2.pred, y_true = df$bad)
# 44.00242
MLmetrics::AUC(y_pred = net2.pred, y_true = df$bad)
# 0.7895033

# LOGISTIC REGRESSION PERFORMANCE
MLmetrics::KS_Stat(y_pred = fitted(mdl2), y_true = df$bad)
# 42.61731
MLmetrics::AUC(y_pred = fitted(mdl2), y_true = df$bad)
# 0.7751298

# MARGINAL EFFECT OF EACH ATTRIBUTE
par(mfrow = c(3, 4))
lapply(1:11, function(i) grnn.margin(net2, i))