I know if we remove some non support vectors, then the discriminant line doesn't change - but by removing some support vectors, does it?
If it does, does the margin width always increase or can it sometimes decrease?
Thanks Rene but I mean if we removed some support vectors and then get a new discriminant line based on remain support vectors, the margin width definitely increases or sometimes can decrease?
I agree with Rene except with the fact that the margin width decreases. Imagine a SVM with the following training data: (0,0) on class -1 and {(1,1), (1,-1)} on class 1. All three vectors become support vectors and the margin (distance between hyperplanes) is 1. If I remove vector (1,-1), the margin increases to sqrt(2). In fact, the margin remains the same or increases, but never decreases, for the reasons given by Rene.
I also agree with Salvador. The margin won't decrease, as this would only occur when points are inserted between the existing support vectors of opposite label.
- Badges
- Science topic