Hi all -
(Re my previous question to the list, I managed to generate an ROC plot
for my SVM by ranking the data using the Decision.Values property.
Thanks for your responses)
I have now started tuning the SVM to get optimal parameters for the RBF
kernel and I ran into a problem. Whatever parameter ranges I specify, I
always get the same error values for all combinations of parameters it
tries. How is "Error" defined in tune.svm? How can I compare this to
results of a single point calculation? Error seems roughly 1- accuracy,
but not exactly.
Thanks for any help.
Ton
=== tuning
> t <- tune(svm, x,y, ranges = list(gamma = 2^(-1:1), cost = 2^3),
control = tune.control(sampling = "cross", performances=TRUE))
> summary(t)
Parameter tuning of `svm':
- sampling method: 10-fold cross validation
- best parameters:
gamma cost
0.5 8
- best performance: 0.1120897
- Detailed performance results:
gamma cost error
1 0.5 8 0.1120897
2 1.0 8 0.1120897
3 2.0 8 0.1120897
=== Single point run
> t <- svm(x,y, cross = 10, gamma = 2^0, cost = 2^3, fitted=TRUE)
> summary(t)
Call:
svm.default(x = x, y = y, gamma = 2^0, cost = 2^3, cross = 10,
fitted = TRUE)
Parameters:
SVM-Type: C-classification
SVM-Kernel: radial
cost: 8
gamma: 1
Number of Support Vectors: 788
( 88 700 )
Number of Classes: 2
Levels:
false true
10-fold cross-validation on training data:
Total Accuracy: 88.91626
Single Accuracies:
82.71605 86.41975 91.35802 91.35802 95.12195 97.53086 90.12346 88.88889
91.35802 74.39024
[[alternative HTML version deleted]]
