Hello,
I have noticed a bug with LSSVM implementation in R. It could be a bug with
the LSSVM itself that causes this problem.
I thought I should post this message to see if anyone else is familiar with
this problem and explain why the result is different for odd and even number
of cases.
Once the hyperplane is found using LSSVM, the prediction results vary when
you predict odd or even number of samples. Why? Here I provide e.g. with
Iris data in R, keep reducing prediction cases one-by-one, you will see the
discrepancy I am talking about. In my own data, this discrepancy between odd
and even number of cases is enhanced by a huge factor.
Thanks,
Parmee
iris <- unique(iris)
rbf <- rbfdot(0.5)
lssvm> k <- kernelMatrix(rbf, as.matrix(iris[,-5]))
lssvm> klir <- lssvm(k, iris[, 5])
lssvm> pre <- predict(klir, k)
> ktest <- as.kernelMatrix(k[1:148,])
> pretest <- predict(klir, ktest)
> table(pretest,iris[1:148,5])
pretest setosa versicolor virginica
setosa 50 0 0
versicolor 0 49 2
virginica 0 1 46
> ktest2 <- as.kernelMatrix(k[1:147,])
> pretest2 <- predict(klir, ktest2)
> table(pretest2,iris[1:147,5])
pretest2 setosa versicolor virginica
setosa 50 0 0
versicolor 0 49 4
virginica 0 1 43
> ktest3 <- as.kernelMatrix(k[1:146,])
> pretest3 <- predict(klir, ktest3)
> table(pretest3,iris[1:146,5])
pretest3 setosa versicolor virginica
setosa 50 0 0
versicolor 0 49 2
virginica 0 1 44
> ktest4 <- as.kernelMatrix(k[1:145,])
> pretest4 <- predict(klir, ktest4)
> table(pretest4,iris[1:145,5])
pretest4 setosa versicolor virginica
setosa 50 0 0
versicolor 0 49 2
virginica 0 1 43
> ktest5 <- as.kernelMatrix(k[1:144,])
> pretest5 <- predict(klir, ktest5)
> table(pretest5,iris[1:144,5])
pretest5 setosa versicolor virginica
setosa 50 0 0
versicolor 0 49 4
virginica 0 1 40
> ktest6 <- as.kernelMatrix(k[1:143,])
> pretest6 <- predict(klir, ktest6)
> table(pretest6,iris[1:143,5])
pretest6 setosa versicolor virginica
setosa 50 0 0
versicolor 0 49 2
virginica 0 1 41
> ktest7 <- as.kernelMatrix(k[1:142,])
> pretest7 <- predict(klir, ktest7)
> table(pretest7,iris[1:142,5])
pretest7 setosa versicolor virginica
setosa 50 0 0
versicolor 0 49 2
virginica 0 1 40
> pretest8 <- predict(klir, ktest8)
> ktest8 <- as.kernelMatrix(k[1:141,])
> table(pretest8,iris[1:141,5])
pretest8 setosa versicolor virginica
setosa 50 0 0
versicolor 0 49 4
virginica 0 1 37
> ktest9 <- as.kernelMatrix(k[1:140,])
> pretest9 <- predict(klir, ktest9)
> table(pretest9,iris[1:140,5])
pretest9 setosa versicolor virginica
setosa 50 0 0
versicolor 0 49 2
virginica 0 1 38
[[alternative HTML version deleted]]
