CPL - Chalmers Publication Library
| Utbildning | Forskning | Styrkeområden | Om Chalmers | In English In English Ej inloggad.

Ridge-Adjusted Slack Variable Optimization for Supervised Classification

Yinan Yu (Institutionen för signaler och system, Signalbehandling) ; Konstantinos I. Diamantaras ; Tomas McKelvey (Institutionen för signaler och system, Signalbehandling) ; S.Y. Kung
IEEE International Workshop on Machine Learning for Signal Processing, Southampton, United Kingdom, September 22-25, 2013 (2161-0363). (2013)
[Konferensbidrag, refereegranskat]

This paper presents an iterative classification algorithm called Ridge-adjusted Slack Variable Optimization (RiSVO). RiSVO is an iterative procedure with two steps: (1) A working subset of the training data is selected so as to reject "extreme" patterns. (2) the decision vector and threshold value are obtained by minimizing the energy function associated with the slack variables. From a computational perspective, we have established a sufficient condition for the "inclusion property" among successive working sets, which allows us to save computation time. Most importantly, under the inclusion property, the monotonic reduction of the energy function can be assured in both substeps at each iteration, thus assuring the convergence of the algorithm. Moreover, ridge regularization is incorporated to improve the robustness and better cope with over-fitting and ill-conditioned problems. To verify the proposed algorithm, we conducted simulations on three data sets from the UCI database: adult, shuttle and bank. Our simulation shows stability and convergence of the RiSVO method. The results also show improvement of performance over the SVM classifier.

Nyckelord: slack energy minimization; kernel method; ridge-regression; classification; training data selection



Denna post skapades 2013-07-12. Senast ändrad 2015-07-03.
CPL Pubid: 180134

 

Läs direkt!

Lokal fulltext (fritt tillgänglig)

Länk till annan sajt (kan kräva inloggning)


Institutioner (Chalmers)

Institutionen för signaler och system, Signalbehandling (1900-2017)

Ämnesområden

Elektroteknik och elektronik

Chalmers infrastruktur