randomForest did not predict serial samples - r

Question

I have data.frame TC, with 17744 observations of 13 variables. The last variable is target: a Factor w/ 2 levels "0", "1".
I do:
n.col <- ncol(TC)
x.train.or <- TC[1:12000, -n.col]
y.train.or <- TC[1:12000, n.col]
x.test.or <- TC[12000:17000, -n.col]
y.test.or <- TC[12000:17000, n.col]
rf.or <- randomForest(y=y.train.or, x=x.train.or, ntree=500, mtry=5,
importance=TRUE, keep.forest=TRUE,
na.action=na.roughfix, replace=FALSE)
pr.or <- predict(rf.or, x.test.or)
table(y.test.or, pr.or, dnn=c("Actual", "Predicted"))
# Predicted
# Actual 0 1
# 0 2424 780
# 1 1056 741
Very bad result.
Then I repeat the model fitting with a random sample:
set.seed <- 123
t.t <- holdout(TC[, n.col], ratio=3/5, mode = "random")
x.train.r <- TC[t.t$tr, - (n.col)]
y.train.r <- TC[t.t$tr, (n.col)]
x.test.r <- TC[t.t$ts, - (n.col)]
rf.r <- randomForest(y=y.train.r, x=x.train.r, ntree=500, mtry=5,
importance=TRUE, keep.forest=TRUE,
na.action=na.roughfix, replace=FALSE)
pr.r <- predict(rf.r, x.test.r)
table(y.test.r, pr.r, dnn=c("Actual", "Predicted"))
# Predicted
# Actual 0 1
# 0 4274 215
# 1 353 2257
Very good result but depended on a way of formation of sample of an one data set.
Problem which I solves assumed only serial sample.
Please, help me!
Answer to questions:
(1)Certainly I do:
library(randomForest)
library(rminer)
(3) I repeat with:
n.col <- ncol(TC)
x.train.or <- TC[1:12000, -n.col]
y.train.or <- TC[1:12000, n.col]
x.test.or <- TC[12001:17000, -n.col]
y.test.or <- TC[12001:17000, n.col]
and receiving the same awful result
Predicted
Actual 0 1
0 2413 790
1 1049 748
(4)There could be a problem in it? Some variables are random on [1:17000], but not random on [1:100]
(I had no rights to drawings).
What to do in that case?

Answer 1

First, it's going to be a little difficult to answer without knowing the state of the data. Sometimes you can be including your test set in your train set if observations repeat themselves in some sort of manner.
One of the best ways to validate your results is through using some sort of cross-validation technique paying heed to making sure you completely separate your test and train set. Below is a good video to watch on that.
http://vimeo.com/75432414