add sensitivity to SVM
Given a training the SVM should be able to tell what inputs are relevant for classification. The sensitivity is defined to be the sum over training samples of the gradient of the output with respect to the input. This is of course kernel dependent. Hence a new function in KernelFunction? must be implemented telling us gradient of the KernelFunction?.
A special case occurs when the kernel is linear, because the gradient is contant and the sum over samples is a waste of time.
Change History (6)
Milestone: |
SVM extension →
later
|
Summary: |
add senitivity to SVM →
add sensitivity to SVM
|
Milestone: |
0.4 →
0.5
|
Status: |
assigned →
new
|
Milestone: |
yat 0.5
|
Resolution: |
→ wontfix
|
Status: |
new →
closed
|
This is not very useful except in the linear case. (see ticket:51)
Input ranking based on a non-linear classifier is not straightforward.