Different ksvm models

ksvm-models.R

> library(kernlab)
> data(iris)
> irismodel1 <- ksvm(Species ~ ., data = iris)
Using automatic sigma estimation (sigest) for RBF or laplace kernel 
> irismodel2 <- ksvm(Species ~ ., data = iris)
Using automatic sigma estimation (sigest) for RBF or laplace kernel 
> table(predict(irismodel1, iris[,-5]), predict(irismodel2, iris[,-5]))
            
             setosa versicolor virginica
  setosa         50          0         0
  versicolor      0         50         0
  virginica       0          1        49
> all.equal(irismodel1, irismodel2)
 [1] "Attributes: < Component 1: Component 1: Numeric: lengths (18, 36) differ >"                                 
 [2] "Attributes: < Component 1: Component 2: Numeric: lengths (25, 39) differ >"                                 
 [3] "Attributes: < Component 1: Component 3: Numeric: lengths (39, 48) differ >"                                 
 [4] "Attributes: < Component 2: Component 1: Numeric: lengths (18, 36) differ >"                                 
 [5] "Attributes: < Component 2: Component 2: Numeric: lengths (25, 39) differ >"                                 
 [6] "Attributes: < Component 2: Component 3: Numeric: lengths (39, 48) differ >"                                 
 [7] "Attributes: < Component 3: Mean relative difference: 0.9913751 >"                                           
 [8] "Attributes: < Component 5: Component 1: Numeric: lengths (18, 36) differ >"                                 
 [9] "Attributes: < Component 5: Component 2: Numeric: lengths (25, 39) differ >"                                 
[10] "Attributes: < Component 5: Component 3: Numeric: lengths (39, 48) differ >"                                 
[11] "Attributes: < Component 7: Mean relative difference: 0.25 >"                                                
[12] "Attributes: < Component 8: 1 string mismatch >"                                                             
[13] "Attributes: < Component 10: target, current do not match when deparsed >"                                   
[14] "Attributes: < Component 15: Mean relative difference: 0.2280702 >"                                          
[15] "Attributes: < Component 16: Mean relative difference: 0.2696004 >"                                          
[16] "Attributes: < Component 21: Numeric: lengths (57, 70) differ >"                                             
[17] "Attributes: < Component 24: Component 1: Attributes: < Component 1: Mean relative difference: 1 > >"        
[18] "Attributes: < Component 24: Component 1: Attributes: < Component 2: Component 1: 18 string mismatches > >"  
[19] "Attributes: < Component 24: Component 1: Numeric: lengths (72, 144) differ >"                               
[20] "Attributes: < Component 24: Component 2: Attributes: < Component 1: Mean relative difference: 0.56 > >"     
[21] "Attributes: < Component 24: Component 2: Attributes: < Component 2: Component 1: 25 string mismatches > >"  
[22] "Attributes: < Component 24: Component 2: Numeric: lengths (100, 156) differ >"                              
[23] "Attributes: < Component 24: Component 3: Attributes: < Component 1: Mean relative difference: 0.2307692 > >"
[24] "Attributes: < Component 24: Component 3: Attributes: < Component 2: Component 1: 39 string mismatches > >"  
[25] "Attributes: < Component 24: Component 3: Numeric: lengths (156, 192) differ >"

This difference is the result of different sigma estimate

irismodel1@kernelf@kpar$sigma
[1] 1.217243
irismodel2@kernelf@kpar$sigma
[1] 0.5681333

Setting sigma manually

irismodel3 <- ksvm(Species ~ ., data = iris, kpar = list(sigma = 1))
irismodel4 <- ksvm(Species ~ ., data = iris, kpar = list(sigma = 1))
all.equal(irismodel3, irismodel4)
[1] TRUE

Looking at the variability of the automatic sigma estimations

sigmas <- replicate(500, ksvm(Species ~ ., data = iris)@kernelf@kpar$sigma)
summary(sigmas)
Min. 1st Qu. Median Mean 3rd Qu. Max.
0.3263 0.6038 0.7315 0.7772 0.8996 1.8320