-
Notifications
You must be signed in to change notification settings - Fork 4
/
Simulation_0530
227 lines (194 loc) · 7.39 KB
/
Simulation_0530
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
#xgb_res=c()
#gbm_res=c()
#for(k in 1:10){
n1=500
n2=500
n=n1+n2
p=50
#beta=1
rho=0.75; #0.75
#generate covaraince matrix V=rho^|i-j|
#we can also generate R then use cholesky to intrdouce cor.
V=matrix(0,ncol=p,nrow=p)
for (i in 1:p) {
for (j in 1:p ){
V[i,j]=rho^abs(i-j)
}
}
X=MASS::mvrnorm(n=n,mu=rep(0,p),Sigma=V)
#X=matrix(runif(n*p,-3,3),nrow=n,byrow=T)
#mu=exp(2*pnorm(X[,10]^2+X[,50]^2-1)) #***
#mu=exp(2*pnorm(X[,30]*X[,10]+X[,50]^2+X[,10]^2-1))
#mu=exp(2*pnorm(sin(X[,10])+X[,50]^2-1))
#mu=exp(2*pnorm(cos(X[,10])+X[,50]^2-1)) #***
#mu=exp(2*pnorm((X[,10]>0.5)+X[,50]^2-1)) #***
#mu=exp(-1+X[,20]^2/0.5+X[,40]^2/0.5+X[,30]^2/0.5+X[,10]^2/0.5+X[,50]^2/0.5)
#mu=2*pnorm(-1+X[,10]*X[,30]+X[,50]^2) # 2*pnorm()
#2.5*pnorm(-1+X[,20]^2/1+X[,50]^2/1+X[,30]*X[,10]+sin(X[,5])+(X[,15]>0)+sin(6*X[,2])+
# 4*X[,4]^3+cos(6*X[,3])+X[,25]^2*X[,45]^2)
mu=0
#for(i in seq(5,p,5)){
for(i in seq(1,p*0.2,1)){
if((i/1)%%4==0){mu=mu+X[,i]^2/0.5}
if((i/1)%%4==1){mu=mu+cos(X[,i])}
if((i/1)%%4==2){mu=mu+sin(i*X[,i])}
if((i/1)%%4==3){mu=mu+X[,i]*X[,i-1]}
}
#mu=exp(-p)*exp(mu) #0.05 FOR 50
#mu=exp((X[,10]>0.8)-1) #*** try this plot(gbm1,c(10,30),best.iter)
mu=0.1*abs(mu)
#T=-(log(runif(n)))/(mu)
#T=(rexp(n,5))/(mu)
T=rweibull(n, shape=2, scale=mu)
#a=2*rbinom(n=n, size=1, prob=1/3); b=runif(n=n,min=0,max=2)
#a[a==0]=b[a==0]
#C=a
C=runif(n, min = 0, max = 2)
C[sample(n,n/3)]=2
obs.time<- pmin(T,C)
status <- T<=C
#table(status)
library(ggplot2)
library("survival")
library('survminer')
library(survcomp)
fit1 <- survfit(Surv(obs.time,status)~1)
plot(fit1)
ggsurvplot(fit1,risk.table = TRUE)
#coxph
require(survival)
df2 <- structure(list(X_c=X[1:n1,], status_c=status[1:n1], obs.time_c=obs.time[1:n1], class = "data.frame"))
fit1=coxph(Surv(obs.time_c, status_c)~ X_c, method="breslow",data=df2)
newdata=structure(list(X_c=X[(n1+1):n,],
.Names = c("X_c"), class = "data.frame"))
cox_pred=predict(fit1,newdata,type="lp")
#gbm cox
library(gbm)
library(survival)
gbm1 <- gbm(Surv(obs.time[1:n1],status[1:n1])~ ., # formula
data=as.data.frame(X[1:n1,]), # dataset
#weights=w,
#var.monotone=c(0,0,0), # -1: monotone decrease, +1: monotone increase, 0: no monotone restrictions
distribution="coxph",
n.trees=2000, # number of trees
shrinkage=0.005, # shrinkage or learning rate, 0.001 to 0.1 usually work
#interaction.depth=1, # 1: additive model, 2: two-way interactions, etc
bag.fraction = 0.5, # subsampling fraction, 0.5 is probably best
train.fraction = 1, # fraction of data for training, first train.fraction*N used for training
cv.folds = 5, # do 5-fold cross-validation
#n.minobsinnode = 10, # minimum total weight needed in each node
keep.data = TRUE,
verbose = TRUE) # print progress
#summary(gbm1)
?gbm
best.iter <- gbm.perf(gbm1,method="cv")
ss=summary(gbm1,n.trees=best.iter) # based on the estimated best number of trees
ss$rel.inf
ss$var
gbm_pred=predict(gbm1,as.data.frame(X[(n1+1):n,]))
survConcordance(Surv(obs.time[(n1+1):n], status[(n1+1):n]) ~ gbm_pred)
library(xgboost)
Dtrain<-xgb.DMatrix(X[1:n1,],label=obs.time[1:n1]*(-(-1)^(as.numeric(status[1:n1]))))
Dtest<-xgb.DMatrix(X[(n1+1):n,],label=obs.time[(n1+1):n]*(-(-1)^(as.numeric(status[(n1+1):n]))))
mylossobj2<-function(preds, dtrain) {
labels <- getinfo(dtrain, "label")
#censor<-attr(dtrain,"censor")
censor= labels>0
labels=abs(labels)
ord<-order(labels)
ran=rank(labels)
d=censor[ord] #status
etas=preds[ord] #linear predictor
haz<-as.numeric(exp(etas)) #w[i]
rsk<-rev(cumsum(rev(haz))) #W[i]
P<-outer (haz,rsk,'/')
P[upper.tri(P)] <- 0
grad<- -(d-P%*%d)
grad=grad[ran]
H1=P
H2=outer(haz^2,rsk^2,'/')
H=H1-H2
H[upper.tri(H)]=0
hess=H%*%d
hess=hess[ran]
return(list(grad = grad, hess = hess))
}
#evalerror2 <- function(preds, dtrain) {
# labels <- getinfo(dtrain, "label")
# censor= labels>0
# labels=abs(labels)
# err <- as.numeric(survConcordance(Surv(labels, censor) ~ preds)$concordance)
# return(list(metric = "cindex",value = -err))
#}
evalerror2 <- function(preds, dtrain) {
labels <- getinfo(dtrain, "label") #labels<-dtrain$label
censor= labels>0
labels=abs(labels)
ord<-order(labels)
d=censor[ord] #status
etas=preds[ord] #linear predictor
haz<-as.numeric(exp(etas)) #w[i]
rsk<-rev(cumsum(rev(haz)))
err <- -2*sum(d*(etas-log(rsk)))/length(labels)
return(list(metric = "cindex",value = err))
}
#xgb_res[k]=as.numeric((survConcordance(Surv(obs.time[(n1+1):n], status[(n1+1):n]) ~ xgb_pred))$concordance)
#gbm_res[k]=as.numeric((survConcordance(Surv(obs.time[(n1+1):n], status[(n1+1):n]) ~ gbm_pred))$concordance)
#survConcordance(Surv(obs.time[(n1+1):n], status[(n1+1):n]) ~ gbm_pred)
#}
#mean(xgb_res)
#mean(gbm_res)
########################### parameter tuning #######################
res=NA
best_param = list()
best_seednumber = 1234
best_loss = Inf
best_loss_index = 0
for (iter in 1:100) {
param <- list(objective = mylossobj2,
eval_metric = evalerror2,
#num_class = 12,
max_depth = sample(3:10, 1),
#eta =runif(1, .01, .1),
eta=runif(1, .001, .01),
gamma = runif(1, 0.0, 0.05),
subsample = 0.5, #runif(1, .5, 1),
#colsample_bytree = runif(1, .5,1),
min_child_weight = sample(1:20, 1),
max_delta_step = sample(1:10, 1),
#colsample_bylevel=runif(1, .5, 1),
lambda=runif(1,0,2),
alpha=runif(1,0,5)
)
cv.nround = 2000
cv.nfold = 5
seed.number = sample.int(10000, 1)[[1]]
set.seed(seed.number)
mdcv <- xgb.cv(data=Dtrain, params = param, nthread=6,
nfold=cv.nfold, nrounds=cv.nround,
verbose = F)
min_loss = min(mdcv$evaluation_log[,'test_cindex_mean'])
min_loss_index = which.min(as.numeric(unlist(mdcv$evaluation_log[,'test_cindex_mean'])))
if (min_loss < best_loss) {
best_loss = min_loss
best_loss_index = min_loss_index
best_seednumber = seed.number
best_param = param
}
print(iter)
}
nround = best_loss_index
set.seed(best_seednumber)
best_param$objective=mylossobj2
md <- xgboost(data=Dtrain, params=best_param, nrounds=nround,nthread=6)
a=xgb.importance(model=md)
xgb_pred=predict(md,Dtest)
#res=list(X,obs.time,status,gbm1,gbm_pred,fit1,cox_pred,a,xgb_pred)
#save(res,file=paste("/home/xw75/zhenyu/", OUTNAME+i, ".Rdata", sep="" ))
survConcordance(Surv(obs.time[(n1+1):n], status[(n1+1):n]) ~ xgb_pred)
survConcordance(Surv(obs.time[(n1+1):n], status[(n1+1):n]) ~ cox_pred)
survConcordance(Surv(obs.time[(n1+1):n], status[(n1+1):n]) ~ gbm_pred)
dd_xgb=data.frame("time"=obs.time[(n1+1):n],"event"=status[(n1+1):n],"score"=(xgb_pred))
sbrier.score2proba(data.tr=dd_xgb, data.ts = dd_xgb, method = "cox")$bsc.integrated
dd_gbm=data.frame("time"=obs.time[(n1+1):n],"event"=status[(n1+1):n],"score"=(gbm_pred))
sbrier.score2proba(data.tr=dd_gbm, data.ts = dd_gbm, method = "cox")$bsc.integrated