[R] SVM cross validation in e1071

[Tao Shi]

Hi list,

Could someone help me to explain why the leave-one-out cross validation results I got from svm using the internal option "cross" are different from those I got manually?  It seems using "cross" to do cross validation, the results are always better.  Please see the code below.  I also include lda as a comparison.

I'm using WinXP, R-2.9.0, and e1071_1.5-19.

Many thanks!

...Tao



##==================================================
## manual
##
##
> set.seed(1234)
> dat <- data.frame( rbind(matrix(rnorm(1000),ncol = 10),  matrix(rnorm(1000, mean=0.6),ncol = 10)))
> cl <- as.factor(rep(1:2, each=100))
> y.lda <- rep(NA, nrow(dat))
> y.svm <- rep(NA, nrow(dat))
> for (i in 1:nrow(dat)){
+ testset <- dat[i, ]
+ trainset <- dat[-i, ]
+     model.lda <- lda(cl[-i]~., data=trainset)
+     model.svm <- svm(cl[-i]~., data=trainset)
+ y.lda[i] <- as.character(predict(model.lda, testset)$class)
+     y.svm[i] <- as.character(predict(model.svm, testset))
+ }
> 
> table(y.lda, cl)
     cl
y.lda  1  2
    1 84 10
    2 16 90
> table(y.svm, cl)
     cl
y.svm  1  2
    1 83  8
    2 17 92
##==========================================
## using internal CV options
##
> z2 <- lda(cl~., data=dat, CV=T)
> table(z2$class, cl)
   cl
     1  2
  1 84 10
  2 16 90
> z <- svm(cl~., data=dat, cross=200)
> table(z$fitted, cl)
   cl
     1  2
  1 93  4
  2  7 96