-
Notifications
You must be signed in to change notification settings - Fork 0
/
08-stacked_model.r
189 lines (147 loc) · 5.95 KB
/
08-stacked_model.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
source("libraries.R")
source("functions.R")
library(gamsel)
library(keras)
source("resources/external/maronna/KurtSDNew.R")
# notas -------------------------------------------------------------------
# VER SI SOBRESAMPLEAR
# parameters --------------------------------------------------------------
semilla = 1993
# load data ----------------------------------------------------
# raw data
x_train_raw = readRDS("data/working/x_train_raw.rds")
x_test_raw = readRDS("data/working/x_test_raw.rds")
# training data for each method
x_train_gam = readRDS("data/working/x_train_gam.rds")
x_train_auto = readRDS("data/working/x_train_auto.rds")
# test data for each method
x_test_gam = readRDS("data/working/x_test_gam.rds")
x_test_auto = readRDS("data/working/x_test_auto.rds")
# target
y_train = readRDS("data/working/y_train.rds")
y_test = readRDS("data/working/y_test.rds")
# scores
scores_train = readRDS("data/working/scores_train.rds")
scores_test = readRDS("data/working/scores_test.rds")
# GAM --------------------------------------------------------------------
# load trained model (GAM with optimal lambda in CV)
cv_mod = readRDS("data/working/gam_cv.rds")
gam_mod = cv_mod$gamsel.fit
i_mod = which(gam_mod$lambdas==cv_mod$lambda.1se)
# get training predictions
pred_train_gam = predict(gam_mod, newdata=as.matrix(x_train_gam), index=i_mod
, type="response") %>% as.vector()
# get test predictions
pred_test_gam = predict(gam_mod, newdata=as.matrix(x_test_gam), index=i_mod
, type="response") %>% as.vector()
# AUTOENCODER ------------------------------------------------------------
# en un contexto de produccion se aplicaria un modelo ya entrenado a cada batch/caso nuevo
# load model (trained only with training)
auto_mod = load_model_hdf5("data/working/model_autoencoder.h5",compile = FALSE)
# get training outlyingness
outl_train_auto = (x_train_auto - predict(auto_mod, x_train_auto))**2 %>% apply(1, mean)
# get test outlyingness
outl_test_auto = (x_test_auto - predict(auto_mod, x_test_auto))**2 %>% apply(1, mean)
# train stacked logistic -------------------------------------------
# COMO EVALUAR SI LOS TRES INDICES MEJORAN PREDICCION:
# si el coeficiente es significativo.
# pero ojo que el test esta sesgado (ver hofling y tibshirani 2007)
# segun hofling y tibshirani hay que entrenar logistic con test data!!!
# (en realidad en varios folds de CV, pero no lo vamos a hacer asi)
# features and logs
dat_test = data.frame(
pred_gam = pred_test_gam
,outl_auto = outl_test_auto
,sofa = scores_test$score_sofa
,sapsii = scores_test$score_sapsiiprob
,oasis = scores_test$score_oasisprob
,mort_7 = y_test
) %>%
mutate_all(list("log"=function(x) ifelse(x==0,0,log(x)))) %>%
select(-mort_7_log)
# formula for model
log_form = mort_7 ~
pred_gam +
outl_auto_log +
sofa +
sapsii +
oasis
# make and save corplot
g_cor_stacked = GGally::ggcorr(dat_test %>% select(labels(terms(log_form)))
,label=T, hjust=1, label_size=4, layout.exp=1
,label_round=2)
ggsave("output/plots/stacked_corplot.png", g_cor_stacked, width=5, height=3)
# train and save binomial glm
log_mod = glm(formula=log_form, family="binomial", data=dat_test)
# save model
saveRDS(log_mod, "output/model_stacked_logistic_test.rds")
# bootstrap
fit_log = function(data) glm(formula=log_form, family="binomial", data=data)
boot = rsample::bootstraps(dat_test, times=500) %>%
mutate(log_mod = map(splits, fit_log)) %>%
mutate(log_res = map(log_mod, broom::tidy))
boot_res = boot %>% unnest(log_res) %>%
group_by(term) %>%
summarise(
"Coef. (media)" = mean(estimate)
,"Des.Est. Coef." = sd(estimate)
,"p-valor (media)" = mean(p.value)
)
# save bootrstap results
saveRDS(boot_res, "output/tables/boot_stacked_logistic_test.rds")
# plots -------------------------------------------------------------------
# DEFINIR SI ESTO TIENE SENTIDO
# table with observed and fitted
dat_p = dat_test %>%
select(labels(terms(log_form)),mort_7) %>%
mutate(
stacked_logistic = predict(log_mod, newdata=dat_test, type="response")
,y = as.factor(mort_7)
) %>%
select(-mort_7)
# ROC curve
library(yardstick)
roc_dat = list(
GAM = roc_curve(dat_p, y, pred_gam)
,SOFA = roc_curve(dat_p, y, sofa)
,SAPSII = roc_curve(dat_p, y, sapsii)
,OASIS = roc_curve(dat_p, y, oasis)
,Autoencoder = roc_curve(dat_p, y, outl_auto_log)
,Stacked_Logistic = roc_curve(dat_p, y, stacked_logistic)
) %>% bind_rows(.id="score")
(
g_roc = ggplot(roc_dat, aes(x=1-specificity, y=sensitivity, color=score)) +
geom_path(cex = 1) +
geom_abline(lty = 3) +
coord_equal() +
theme_bw() +
NULL
)
ggsave("output/plots/roc_stacked.png", g_roc, width=6, height=6)
# tables ------------------------------------------------------------------
# VER SI ESTO TIENE SENTIDO
# AUROC
tab_auroc = dat_p %>%
pivot_longer(names_to="score", values_to="value", -y) %>%
split(.$score) %>%
map_dfc(function(d) roc_auc(d, truth=y, value)$.estimate)
saveRDS(tab_auroc, "output/tables/auroc_stacked.rds")
# OLD ---------------------------------------------------------------------
# KSD (se saca porque es dificil de aplicar en casos nuevos)
# x_train_ksd = readRDS("data/working/x_train_ksd.rds")
# x_test_ksd = readRDS("data/working/x_test_ksd.rds")
# # KSD ------------------------------------------------------------
#
# # se ejecuta para todos los datos (asi funcionaria en un contexto de produccion)
# x_ksd = bind_rows(
# bind_cols(id_tot = x_train_raw$id_tot, x_train_ksd)
# ,bind_cols(id_tot = x_test_raw$id_tot, x_test_ksd)
# )
# # apply KSD
# ksd_mod = KurtSDNew(X = x_ksd %>% select(-id_tot))
# out_ksd = tibble(id_tot = x_ksd$id_tot, out = ksd_mod$tl[[1]])
#
# # get training outlyingness
# outl_train_ksd = out_ksd %>% dplyr::filter(id_tot %in% x_train_raw$id_tot) %>% pull(out)
# # get test outlyingness
# outl_test_ksd = out_ksd %>% dplyr::filter(id_tot %in% x_test_raw$id_tot) %>% pull(out)