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
|
validate.tree <- function(fit, method, B, bw, rule, type, sls, aics,
pr=TRUE, k, rand, xval = 10,
FUN, ...) {
wm <- if(inherits(fit,'tree'))'tree' else
if(inherits(fit,'rpart'))'rpart' else 'none'
if(wm=='none') stop("Not legitimate tree")
# if(.R. && wm=='rpart') require('maptree') # for prune.Rpart
if(missing(FUN))
FUN <- switch(wm,
tree=prune.tree,
rpart=function(...,k)prune.rpart(...,cp=k))
act <- (fit$call)$na.action
if(!length(act))
act <- function(x) x
m <- model.frame(fit, na.action = act)
if(!is.data.frame(m)) stop('you must specify model=T in the fit')
y <- model.extract(m, response)
binary <- is.logical(y) ||
((length(un <- sort(unique(y[!is.na(y)]))) ==
2) && un[1] == 0 && un[2] == 1)
if(binary && is.factor(y)) y <- as.numeric(y) - 1 ## 12dec02
call <- match.call()
method <- call$method
size <- NULL
if(missing(k)) {
if(wm=='tree') {
ff <- FUN(fit, ...)
k <- ff$k
size <- ff$size[is.finite(k)]
k <- k[is.finite(k)] # tree makes first k -Inf 7dec02
} else {
k <- fit$cptable[,'CP']
size <- fit$cptable[,'nsplit']
}
}
if(missing(rand))
rand <- sample(xval, length(m[[1]]), replace = TRUE)
which <- unique(rand)
pdyx.app <- pdyx.val <- pb.app <- pb.val <- double(length(k))
l <- 0
for(kk in k) {
l <- l + 1
dyx.val <- dyx.app <- b.val <- b.app <- double(length(which))
j <- 0
for(i in which) {
j <- j + 1
s <- rand != i
tlearn <- switch(wm,tree=tree(model=m[s,]),rpart=rpart(model=m[s,]))
papp <- if(kk == 0) tlearn else FUN(tlearn, k = kk, ...)
if(nrow(papp$frame) == 1) {
dyx.app[j] <- dyx.val[j] <- 0 #no splits
b.app[j] <- b.val[j] <- mean((y - mean(y))^2, na.rm = TRUE)
}
else {
yhat <- predict(papp, newdata = m[s, ])
if(is.matrix(yhat) && ncol(yhat) > 1)
yhat <- yhat[,ncol(yhat),drop=TRUE]
## tree with factor binary y 7dec02
b.app[j] <- mean((yhat - y[s])^2)
dyx.app[j] <- if(binary) somers2(yhat, y[s])["Dxy"] else
rcorr.cens(yhat, y[s])["Dxy"]
s <- rand == i
yhat <- predict(papp, newdata = m[s, ])
b.val[j] <- mean((yhat - y[s])^2)
dyx.val[j] <- if(binary) somers2(yhat, y[s])["Dxy"] else
rcorr.cens(yhat, y[s])["Dxy"]
}
}
pdyx.app[l] <- mean(dyx.app)
pdyx.val[l] <- mean(dyx.val)
pb.app[l] <- mean(b.app)
pb.val[l] <- mean(b.val)
if(pr) {
dyx.app <- c(dyx.app, pdyx.app[l])
dyx.val <- c(dyx.val, pdyx.val[l])
b.app <- c(b.app, pb.app[l])
b.val <- c(b.val, pb.val[l])
cat("\n\nk=", format(kk), ":\n\n")
dyx <- cbind(dyx.app, dyx.val, b.app, b.val)
dimnames(dyx) <- list(c(as.character(1:j), "Mean"),
c("Dxy Training", "Dxy Test", "MSE Training",
"MSE Test"))
print(dyx)
}
}
structure(list(k = k, size = size, dxy.app = pdyx.app, dxy.val =
pdyx.val, mse.app = pb.app, mse.val = pb.val,
binary = binary, xval = xval),
class = "validate.tree")
}
print.validate.tree <- function(x, ...) {
cat(x$xval, "-fold cross-validation\n\n", sep = "")
w <- cbind(k = x$k, size = x$size, Dxy.apparent = x$dxy.app,
Dxy.val = x$dxy.val, MSE.apparent = x$mse.app, MSE.val =
x$mse.val)
if(x$binary)
dimnames(w) <- list(NULL, c("k", if(length(x$size)) "size",
"Dxy.apparent", "Dxy.val", "Brier.apparent",
"Brier.val"))
invisible(print(w))
}
plot.validate.tree <- function(x, what = c("mse", "dxy"),
legendloc = locator, ...) {
obj <- x
if(length(obj$size)) {
x <- obj$size
xlab <- "Number of Nodes"
}
else {
x <- obj$k
xlab <- "Cost/Complexity Parameter"
}
if("mse" %in% what) {
blab <- if(obj$binary) "Brier Score" else "Mean Squared Error"
ylim <- range(c(obj$mse.app, obj$mse.val))
plot(x, obj$mse.app, xlab = xlab, ylab = blab, ylim = ylim,
type = "n")
lines(x, obj$mse.app, lty = 3)
lines(x, obj$mse.val, lty = 1)
title(sub = paste(obj$xval, "-fold cross-validation", sep = ""),
adj = 0)
if(is.function(legendloc))
legend(legendloc(1), c("Apparent", "Cross-validated"),
lty = c(3, 1), bty = "n") else {
par(usr=c(0,1,0,1))
legend(legendloc[1],legendloc[2],
c("Apparent", "Cross-validated"),
lty = c(3, 1), bty = "n")
}
}
if("dxy" %in% what) {
ylim <- range(c(obj$dxy.app, obj$dxy.val))
plot(x, obj$dxy.app, xlab = xlab, ylab = "Somers' Dxy", ylim =
ylim, type = "n")
lines(x, obj$dxy.app, lty = 3)
lines(x, obj$dxy.val, lty = 1)
title(sub = paste(obj$xval, "-fold cross-validation", sep = ""),
adj = 0)
if(is.function(legendloc))
legend(legendloc(1), c("Apparent", "Cross-validated"),
lty = c(3, 1), bty = "n") else {
par(usr=c(0,1,0,1))
legend(legendloc[1],legendloc[2],
c("Apparent", "Cross-validated"),
lty = c(3, 1), bty = "n")
}
}
invisible()
}
validate.rpart <- function(fit, ...) validate.tree(fit, ...)
|
