Multiclass classification by L1 norm Support Vector Machine
Sandor Szedmak, John Shawe-Taylor, Craig .J. Saunders and David .R. Hardoon
In: Pattern Recognition and Machine Learning in Computer Vision Workshop, 02-04 May 2004, Grenoble, France.

Abstract

The multiclass classification attracts a lot of attention in recent time. It has no such an elaborated theoretical foundation than the binary classification does. Rifkin et al. (2004) \cite{Rifkin2004} collects the known approaches to show the One-Vs-All approach may work sufficiently well comparing to all others. We present an extension of the One-Vs-All framework based on $L_{1}$ norm Support Vector Machine which includes the simultaneous optimisation of the different margins. We show an algorithm based on Danztig-Wolfe decomposition which holds the computational complexity of the problem on the same level that the One-Vs-All gives.

PDF - Requires Adobe Acrobat Reader or other PDF viewer.

[Edit]