Abstract
We propose the Bouligand influence function (BIF) as a
new concept for robust statistics. The BIF is a modification of F.R.
Hampel's influence function (IF) and is based on a special cone
derivative instead of the usual G{\^a}teaux-derivative. If the BIF
does exist, then the IF does also exist and both are equal. The
usefulness of Bouligand-derivatives to robust statistics is
explained.
In the second part of the talk we apply the BIF to support vector
machines based on a non-smooth loss function for which the influence
was unknown. We show for the regression case that many support
vector machines based on a Lipschitz continuous loss function and a
bounded kernel have a bounded BIF and hence also have a bounded IF.
In this respect such SVMs are therefore robust. Special cases are
SVMs based on the $\epsilon$-insensitive loss, Huber's loss, and
kernel based quantile regression based on the pinball loss.
new concept for robust statistics. The BIF is a modification of F.R.
Hampel's influence function (IF) and is based on a special cone
derivative instead of the usual G{\^a}teaux-derivative. If the BIF
does exist, then the IF does also exist and both are equal. The
usefulness of Bouligand-derivatives to robust statistics is
explained.
In the second part of the talk we apply the BIF to support vector
machines based on a non-smooth loss function for which the influence
was unknown. We show for the regression case that many support
vector machines based on a Lipschitz continuous loss function and a
bounded kernel have a bounded BIF and hence also have a bounded IF.
In this respect such SVMs are therefore robust. Special cases are
SVMs based on the $\epsilon$-insensitive loss, Huber's loss, and
kernel based quantile regression based on the pinball loss.
Original language | English |
---|---|
Title of host publication | 15th Annual Meeting of the Belgian Statistical Society, Oct 18-20, 2007, Antwerp (Belgium) |
Publication status | Published - 18 Oct 2007 |
Publication series
Name | 15th Annual Meeting of the Belgian Statistical Society, Oct 18-20, 2007, Antwerp (Belgium) |
---|
Keywords
- SVM
- robustness
- Bouligand