Multiclass classification by l1 norm support vector machine

Multiclass classification by L1 norm Support Vector Machine Page 1 Multiclass classification by L1 norm Support Vector Machine Sandor Szedmak, joint work with John Shawe-Taylor, Craig J. Saunders and David R. Hardoon ISIS Group School of Electronics and Computer Science University of Southampton PASCAL Project June 14, 2004 Page 2 Contents • A (short) introduction • Minimax framework on One-Vs-All • How to solve it? – L1 norm SVM – Minimax formulation – The framework of the solution – The “trick” 1 Page 3 A (short) introduction Multiclass classification: • There are several approach, but no good “idea”,theoretical foundation • Approaches: – Separated subproblems, eg One-Vs-All, All-Vs-All, – Model based solution for the full problem, eg CCA(KCCA), • Difficulties on large scale problems(>>1000 cases, >>1000 features) Our objective to find a solution framework for the full problem with enormous size…