Quadric Multi-Class Support Vector Machines

• Activity Number: 551
• Type: Topic Contributed
• Date/Time: Wednesday, August 6, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Nicolas Wicker*+ and Yann Guermeur
• Companies: Universite Lille and CNRS
• Keywords: Support Vector Machines ; Multi-class Support Vector Machines ; quadrics ; maximum margin hyperplane ; SVM ; MSVM

Abstract

Support vector machines (SVM) are today among the best and most used methods for supervised learning. They are particularly suitable when dealing with a two-class problem as they separate linearly points using a maximum margin hyperplane. Unfortunately, this concept does not extend nicely to the multi-class case. This explains why among the many models proposed for multi-class SVM (MSVM), none is clearly favored. We propose to extend SVM by replacing the separating hyperplanes by quadrics following the tracks of support vector clustering. This presents the advantage of giving a clear geometric interpretation of MSVM and of enabling the identification of outliers.